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 and msb counts don't match, this interrupted a write */
648 if (*cnt != rb_time_cnt(msb))
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);
703 static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
705 unsigned long cnt, top, bottom, msb;
706 unsigned long cnt2, top2, bottom2, msb2;
709 /* The cmpxchg always fails if it interrupted an update */
710 if (!__rb_time_read(t, &val, &cnt2))
716 cnt = local_read(&t->cnt);
717 if ((cnt & 3) != cnt2)
722 rb_time_split(val, &top, &bottom, &msb);
723 top = rb_time_val_cnt(top, cnt);
724 bottom = rb_time_val_cnt(bottom, cnt);
726 rb_time_split(set, &top2, &bottom2, &msb2);
727 top2 = rb_time_val_cnt(top2, cnt2);
728 bottom2 = rb_time_val_cnt(bottom2, cnt2);
730 if (!rb_time_read_cmpxchg(&t->cnt, cnt, cnt2))
732 if (!rb_time_read_cmpxchg(&t->msb, msb, msb2))
734 if (!rb_time_read_cmpxchg(&t->top, top, top2))
736 if (!rb_time_read_cmpxchg(&t->bottom, bottom, bottom2))
743 /* local64_t always succeeds */
745 static inline bool rb_time_read(rb_time_t *t, u64 *ret)
747 *ret = local64_read(&t->time);
750 static void rb_time_set(rb_time_t *t, u64 val)
752 local64_set(&t->time, val);
755 static bool rb_time_cmpxchg(rb_time_t *t, u64 expect, u64 set)
757 return local64_try_cmpxchg(&t->time, &expect, set);
762 * Enable this to make sure that the event passed to
763 * ring_buffer_event_time_stamp() is not committed and also
764 * is on the buffer that it passed in.
766 //#define RB_VERIFY_EVENT
767 #ifdef RB_VERIFY_EVENT
768 static struct list_head *rb_list_head(struct list_head *list);
769 static void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
772 struct buffer_page *page = cpu_buffer->commit_page;
773 struct buffer_page *tail_page = READ_ONCE(cpu_buffer->tail_page);
774 struct list_head *next;
776 unsigned long addr = (unsigned long)event;
780 /* Make sure the event exists and is not committed yet */
782 if (page == tail_page || WARN_ON_ONCE(stop++ > 100))
784 commit = local_read(&page->page->commit);
785 write = local_read(&page->write);
786 if (addr >= (unsigned long)&page->page->data[commit] &&
787 addr < (unsigned long)&page->page->data[write])
790 next = rb_list_head(page->list.next);
791 page = list_entry(next, struct buffer_page, list);
796 static inline void verify_event(struct ring_buffer_per_cpu *cpu_buffer,
803 * The absolute time stamp drops the 5 MSBs and some clocks may
804 * require them. The rb_fix_abs_ts() will take a previous full
805 * time stamp, and add the 5 MSB of that time stamp on to the
806 * saved absolute time stamp. Then they are compared in case of
807 * the unlikely event that the latest time stamp incremented
810 static inline u64 rb_fix_abs_ts(u64 abs, u64 save_ts)
812 if (save_ts & TS_MSB) {
813 abs |= save_ts & TS_MSB;
814 /* Check for overflow */
815 if (unlikely(abs < save_ts))
821 static inline u64 rb_time_stamp(struct trace_buffer *buffer);
824 * ring_buffer_event_time_stamp - return the event's current time stamp
825 * @buffer: The buffer that the event is on
826 * @event: the event to get the time stamp of
828 * Note, this must be called after @event is reserved, and before it is
829 * committed to the ring buffer. And must be called from the same
830 * context where the event was reserved (normal, softirq, irq, etc).
832 * Returns the time stamp associated with the current event.
833 * If the event has an extended time stamp, then that is used as
834 * the time stamp to return.
835 * In the highly unlikely case that the event was nested more than
836 * the max nesting, then the write_stamp of the buffer is returned,
837 * otherwise current time is returned, but that really neither of
838 * the last two cases should ever happen.
840 u64 ring_buffer_event_time_stamp(struct trace_buffer *buffer,
841 struct ring_buffer_event *event)
843 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[smp_processor_id()];
847 /* If the event includes an absolute time, then just use that */
848 if (event->type_len == RINGBUF_TYPE_TIME_STAMP) {
849 ts = rb_event_time_stamp(event);
850 return rb_fix_abs_ts(ts, cpu_buffer->tail_page->page->time_stamp);
853 nest = local_read(&cpu_buffer->committing);
854 verify_event(cpu_buffer, event);
855 if (WARN_ON_ONCE(!nest))
858 /* Read the current saved nesting level time stamp */
859 if (likely(--nest < MAX_NEST))
860 return cpu_buffer->event_stamp[nest];
862 /* Shouldn't happen, warn if it does */
863 WARN_ONCE(1, "nest (%d) greater than max", nest);
866 /* Can only fail on 32 bit */
867 if (!rb_time_read(&cpu_buffer->write_stamp, &ts))
868 /* Screw it, just read the current time */
869 ts = rb_time_stamp(cpu_buffer->buffer);
875 * ring_buffer_nr_pages - get the number of buffer pages in the ring buffer
876 * @buffer: The ring_buffer to get the number of pages from
877 * @cpu: The cpu of the ring_buffer to get the number of pages from
879 * Returns the number of pages used by a per_cpu buffer of the ring buffer.
881 size_t ring_buffer_nr_pages(struct trace_buffer *buffer, int cpu)
883 return buffer->buffers[cpu]->nr_pages;
887 * ring_buffer_nr_dirty_pages - get the number of used pages in the ring buffer
888 * @buffer: The ring_buffer to get the number of pages from
889 * @cpu: The cpu of the ring_buffer to get the number of pages from
891 * Returns the number of pages that have content in the ring buffer.
893 size_t ring_buffer_nr_dirty_pages(struct trace_buffer *buffer, int cpu)
899 read = local_read(&buffer->buffers[cpu]->pages_read);
900 lost = local_read(&buffer->buffers[cpu]->pages_lost);
901 cnt = local_read(&buffer->buffers[cpu]->pages_touched);
903 if (WARN_ON_ONCE(cnt < lost))
908 /* The reader can read an empty page, but not more than that */
910 WARN_ON_ONCE(read > cnt + 1);
917 static __always_inline bool full_hit(struct trace_buffer *buffer, int cpu, int full)
919 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
923 nr_pages = cpu_buffer->nr_pages;
924 if (!nr_pages || !full)
927 dirty = ring_buffer_nr_dirty_pages(buffer, cpu);
929 return (dirty * 100) > (full * nr_pages);
933 * rb_wake_up_waiters - wake up tasks waiting for ring buffer input
935 * Schedules a delayed work to wake up any task that is blocked on the
936 * ring buffer waiters queue.
938 static void rb_wake_up_waiters(struct irq_work *work)
940 struct rb_irq_work *rbwork = container_of(work, struct rb_irq_work, work);
942 wake_up_all(&rbwork->waiters);
943 if (rbwork->full_waiters_pending || rbwork->wakeup_full) {
944 rbwork->wakeup_full = false;
945 rbwork->full_waiters_pending = false;
946 wake_up_all(&rbwork->full_waiters);
951 * ring_buffer_wake_waiters - wake up any waiters on this ring buffer
952 * @buffer: The ring buffer to wake waiters on
953 * @cpu: The CPU buffer to wake waiters on
955 * In the case of a file that represents a ring buffer is closing,
956 * it is prudent to wake up any waiters that are on this.
958 void ring_buffer_wake_waiters(struct trace_buffer *buffer, int cpu)
960 struct ring_buffer_per_cpu *cpu_buffer;
961 struct rb_irq_work *rbwork;
966 if (cpu == RING_BUFFER_ALL_CPUS) {
968 /* Wake up individual ones too. One level recursion */
969 for_each_buffer_cpu(buffer, cpu)
970 ring_buffer_wake_waiters(buffer, cpu);
972 rbwork = &buffer->irq_work;
974 if (WARN_ON_ONCE(!buffer->buffers))
976 if (WARN_ON_ONCE(cpu >= nr_cpu_ids))
979 cpu_buffer = buffer->buffers[cpu];
980 /* The CPU buffer may not have been initialized yet */
983 rbwork = &cpu_buffer->irq_work;
986 rbwork->wait_index++;
987 /* make sure the waiters see the new index */
990 rb_wake_up_waiters(&rbwork->work);
994 * ring_buffer_wait - wait for input to the ring buffer
995 * @buffer: buffer to wait on
996 * @cpu: the cpu buffer to wait on
997 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
999 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1000 * as data is added to any of the @buffer's cpu buffers. Otherwise
1001 * it will wait for data to be added to a specific cpu buffer.
1003 int ring_buffer_wait(struct trace_buffer *buffer, int cpu, int full)
1005 struct ring_buffer_per_cpu *cpu_buffer;
1007 struct rb_irq_work *work;
1012 * Depending on what the caller is waiting for, either any
1013 * data in any cpu buffer, or a specific buffer, put the
1014 * caller on the appropriate wait queue.
1016 if (cpu == RING_BUFFER_ALL_CPUS) {
1017 work = &buffer->irq_work;
1018 /* Full only makes sense on per cpu reads */
1021 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1023 cpu_buffer = buffer->buffers[cpu];
1024 work = &cpu_buffer->irq_work;
1027 wait_index = READ_ONCE(work->wait_index);
1031 prepare_to_wait(&work->full_waiters, &wait, TASK_INTERRUPTIBLE);
1033 prepare_to_wait(&work->waiters, &wait, TASK_INTERRUPTIBLE);
1036 * The events can happen in critical sections where
1037 * checking a work queue can cause deadlocks.
1038 * After adding a task to the queue, this flag is set
1039 * only to notify events to try to wake up the queue
1042 * We don't clear it even if the buffer is no longer
1043 * empty. The flag only causes the next event to run
1044 * irq_work to do the work queue wake up. The worse
1045 * that can happen if we race with !trace_empty() is that
1046 * an event will cause an irq_work to try to wake up
1049 * There's no reason to protect this flag either, as
1050 * the work queue and irq_work logic will do the necessary
1051 * synchronization for the wake ups. The only thing
1052 * that is necessary is that the wake up happens after
1053 * a task has been queued. It's OK for spurious wake ups.
1056 work->full_waiters_pending = true;
1058 work->waiters_pending = true;
1060 if (signal_pending(current)) {
1065 if (cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer))
1068 if (cpu != RING_BUFFER_ALL_CPUS &&
1069 !ring_buffer_empty_cpu(buffer, cpu)) {
1070 unsigned long flags;
1077 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1078 pagebusy = cpu_buffer->reader_page == cpu_buffer->commit_page;
1079 done = !pagebusy && full_hit(buffer, cpu, full);
1081 if (!cpu_buffer->shortest_full ||
1082 cpu_buffer->shortest_full > full)
1083 cpu_buffer->shortest_full = full;
1084 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1091 /* Make sure to see the new wait index */
1093 if (wait_index != work->wait_index)
1098 finish_wait(&work->full_waiters, &wait);
1100 finish_wait(&work->waiters, &wait);
1106 * ring_buffer_poll_wait - poll on buffer input
1107 * @buffer: buffer to wait on
1108 * @cpu: the cpu buffer to wait on
1109 * @filp: the file descriptor
1110 * @poll_table: The poll descriptor
1111 * @full: wait until the percentage of pages are available, if @cpu != RING_BUFFER_ALL_CPUS
1113 * If @cpu == RING_BUFFER_ALL_CPUS then the task will wake up as soon
1114 * as data is added to any of the @buffer's cpu buffers. Otherwise
1115 * it will wait for data to be added to a specific cpu buffer.
1117 * Returns EPOLLIN | EPOLLRDNORM if data exists in the buffers,
1120 __poll_t ring_buffer_poll_wait(struct trace_buffer *buffer, int cpu,
1121 struct file *filp, poll_table *poll_table, int full)
1123 struct ring_buffer_per_cpu *cpu_buffer;
1124 struct rb_irq_work *work;
1126 if (cpu == RING_BUFFER_ALL_CPUS) {
1127 work = &buffer->irq_work;
1130 if (!cpumask_test_cpu(cpu, buffer->cpumask))
1133 cpu_buffer = buffer->buffers[cpu];
1134 work = &cpu_buffer->irq_work;
1138 poll_wait(filp, &work->full_waiters, poll_table);
1139 work->full_waiters_pending = true;
1140 if (!cpu_buffer->shortest_full ||
1141 cpu_buffer->shortest_full > full)
1142 cpu_buffer->shortest_full = full;
1144 poll_wait(filp, &work->waiters, poll_table);
1145 work->waiters_pending = true;
1149 * There's a tight race between setting the waiters_pending and
1150 * checking if the ring buffer is empty. Once the waiters_pending bit
1151 * is set, the next event will wake the task up, but we can get stuck
1152 * if there's only a single event in.
1154 * FIXME: Ideally, we need a memory barrier on the writer side as well,
1155 * but adding a memory barrier to all events will cause too much of a
1156 * performance hit in the fast path. We only need a memory barrier when
1157 * the buffer goes from empty to having content. But as this race is
1158 * extremely small, and it's not a problem if another event comes in, we
1159 * will fix it later.
1164 return full_hit(buffer, cpu, full) ? EPOLLIN | EPOLLRDNORM : 0;
1166 if ((cpu == RING_BUFFER_ALL_CPUS && !ring_buffer_empty(buffer)) ||
1167 (cpu != RING_BUFFER_ALL_CPUS && !ring_buffer_empty_cpu(buffer, cpu)))
1168 return EPOLLIN | EPOLLRDNORM;
1172 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
1173 #define RB_WARN_ON(b, cond) \
1175 int _____ret = unlikely(cond); \
1177 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
1178 struct ring_buffer_per_cpu *__b = \
1180 atomic_inc(&__b->buffer->record_disabled); \
1182 atomic_inc(&b->record_disabled); \
1188 /* Up this if you want to test the TIME_EXTENTS and normalization */
1189 #define DEBUG_SHIFT 0
1191 static inline u64 rb_time_stamp(struct trace_buffer *buffer)
1195 /* Skip retpolines :-( */
1196 if (IS_ENABLED(CONFIG_RETPOLINE) && likely(buffer->clock == trace_clock_local))
1197 ts = trace_clock_local();
1199 ts = buffer->clock();
1201 /* shift to debug/test normalization and TIME_EXTENTS */
1202 return ts << DEBUG_SHIFT;
1205 u64 ring_buffer_time_stamp(struct trace_buffer *buffer)
1209 preempt_disable_notrace();
1210 time = rb_time_stamp(buffer);
1211 preempt_enable_notrace();
1215 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
1217 void ring_buffer_normalize_time_stamp(struct trace_buffer *buffer,
1220 /* Just stupid testing the normalize function and deltas */
1221 *ts >>= DEBUG_SHIFT;
1223 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
1226 * Making the ring buffer lockless makes things tricky.
1227 * Although writes only happen on the CPU that they are on,
1228 * and they only need to worry about interrupts. Reads can
1229 * happen on any CPU.
1231 * The reader page is always off the ring buffer, but when the
1232 * reader finishes with a page, it needs to swap its page with
1233 * a new one from the buffer. The reader needs to take from
1234 * the head (writes go to the tail). But if a writer is in overwrite
1235 * mode and wraps, it must push the head page forward.
1237 * Here lies the problem.
1239 * The reader must be careful to replace only the head page, and
1240 * not another one. As described at the top of the file in the
1241 * ASCII art, the reader sets its old page to point to the next
1242 * page after head. It then sets the page after head to point to
1243 * the old reader page. But if the writer moves the head page
1244 * during this operation, the reader could end up with the tail.
1246 * We use cmpxchg to help prevent this race. We also do something
1247 * special with the page before head. We set the LSB to 1.
1249 * When the writer must push the page forward, it will clear the
1250 * bit that points to the head page, move the head, and then set
1251 * the bit that points to the new head page.
1253 * We also don't want an interrupt coming in and moving the head
1254 * page on another writer. Thus we use the second LSB to catch
1257 * head->list->prev->next bit 1 bit 0
1260 * Points to head page 0 1
1263 * Note we can not trust the prev pointer of the head page, because:
1265 * +----+ +-----+ +-----+
1266 * | |------>| T |---X--->| N |
1268 * +----+ +-----+ +-----+
1271 * +----------| R |----------+ |
1275 * Key: ---X--> HEAD flag set in pointer
1280 * (see __rb_reserve_next() to see where this happens)
1282 * What the above shows is that the reader just swapped out
1283 * the reader page with a page in the buffer, but before it
1284 * could make the new header point back to the new page added
1285 * it was preempted by a writer. The writer moved forward onto
1286 * the new page added by the reader and is about to move forward
1289 * You can see, it is legitimate for the previous pointer of
1290 * the head (or any page) not to point back to itself. But only
1294 #define RB_PAGE_NORMAL 0UL
1295 #define RB_PAGE_HEAD 1UL
1296 #define RB_PAGE_UPDATE 2UL
1299 #define RB_FLAG_MASK 3UL
1301 /* PAGE_MOVED is not part of the mask */
1302 #define RB_PAGE_MOVED 4UL
1305 * rb_list_head - remove any bit
1307 static struct list_head *rb_list_head(struct list_head *list)
1309 unsigned long val = (unsigned long)list;
1311 return (struct list_head *)(val & ~RB_FLAG_MASK);
1315 * rb_is_head_page - test if the given page is the head page
1317 * Because the reader may move the head_page pointer, we can
1318 * not trust what the head page is (it may be pointing to
1319 * the reader page). But if the next page is a header page,
1320 * its flags will be non zero.
1323 rb_is_head_page(struct buffer_page *page, struct list_head *list)
1327 val = (unsigned long)list->next;
1329 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
1330 return RB_PAGE_MOVED;
1332 return val & RB_FLAG_MASK;
1338 * The unique thing about the reader page, is that, if the
1339 * writer is ever on it, the previous pointer never points
1340 * back to the reader page.
1342 static bool rb_is_reader_page(struct buffer_page *page)
1344 struct list_head *list = page->list.prev;
1346 return rb_list_head(list->next) != &page->list;
1350 * rb_set_list_to_head - set a list_head to be pointing to head.
1352 static void rb_set_list_to_head(struct list_head *list)
1356 ptr = (unsigned long *)&list->next;
1357 *ptr |= RB_PAGE_HEAD;
1358 *ptr &= ~RB_PAGE_UPDATE;
1362 * rb_head_page_activate - sets up head page
1364 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
1366 struct buffer_page *head;
1368 head = cpu_buffer->head_page;
1373 * Set the previous list pointer to have the HEAD flag.
1375 rb_set_list_to_head(head->list.prev);
1378 static void rb_list_head_clear(struct list_head *list)
1380 unsigned long *ptr = (unsigned long *)&list->next;
1382 *ptr &= ~RB_FLAG_MASK;
1386 * rb_head_page_deactivate - clears head page ptr (for free list)
1389 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
1391 struct list_head *hd;
1393 /* Go through the whole list and clear any pointers found. */
1394 rb_list_head_clear(cpu_buffer->pages);
1396 list_for_each(hd, cpu_buffer->pages)
1397 rb_list_head_clear(hd);
1400 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
1401 struct buffer_page *head,
1402 struct buffer_page *prev,
1403 int old_flag, int new_flag)
1405 struct list_head *list;
1406 unsigned long val = (unsigned long)&head->list;
1411 val &= ~RB_FLAG_MASK;
1413 ret = cmpxchg((unsigned long *)&list->next,
1414 val | old_flag, val | new_flag);
1416 /* check if the reader took the page */
1417 if ((ret & ~RB_FLAG_MASK) != val)
1418 return RB_PAGE_MOVED;
1420 return ret & RB_FLAG_MASK;
1423 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
1424 struct buffer_page *head,
1425 struct buffer_page *prev,
1428 return rb_head_page_set(cpu_buffer, head, prev,
1429 old_flag, RB_PAGE_UPDATE);
1432 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
1433 struct buffer_page *head,
1434 struct buffer_page *prev,
1437 return rb_head_page_set(cpu_buffer, head, prev,
1438 old_flag, RB_PAGE_HEAD);
1441 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
1442 struct buffer_page *head,
1443 struct buffer_page *prev,
1446 return rb_head_page_set(cpu_buffer, head, prev,
1447 old_flag, RB_PAGE_NORMAL);
1450 static inline void rb_inc_page(struct buffer_page **bpage)
1452 struct list_head *p = rb_list_head((*bpage)->list.next);
1454 *bpage = list_entry(p, struct buffer_page, list);
1457 static struct buffer_page *
1458 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
1460 struct buffer_page *head;
1461 struct buffer_page *page;
1462 struct list_head *list;
1465 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
1469 list = cpu_buffer->pages;
1470 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
1473 page = head = cpu_buffer->head_page;
1475 * It is possible that the writer moves the header behind
1476 * where we started, and we miss in one loop.
1477 * A second loop should grab the header, but we'll do
1478 * three loops just because I'm paranoid.
1480 for (i = 0; i < 3; i++) {
1482 if (rb_is_head_page(page, page->list.prev)) {
1483 cpu_buffer->head_page = page;
1487 } while (page != head);
1490 RB_WARN_ON(cpu_buffer, 1);
1495 static bool rb_head_page_replace(struct buffer_page *old,
1496 struct buffer_page *new)
1498 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
1501 val = *ptr & ~RB_FLAG_MASK;
1502 val |= RB_PAGE_HEAD;
1504 return try_cmpxchg(ptr, &val, (unsigned long)&new->list);
1508 * rb_tail_page_update - move the tail page forward
1510 static void rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
1511 struct buffer_page *tail_page,
1512 struct buffer_page *next_page)
1514 unsigned long old_entries;
1515 unsigned long old_write;
1518 * The tail page now needs to be moved forward.
1520 * We need to reset the tail page, but without messing
1521 * with possible erasing of data brought in by interrupts
1522 * that have moved the tail page and are currently on it.
1524 * We add a counter to the write field to denote this.
1526 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
1527 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
1529 local_inc(&cpu_buffer->pages_touched);
1531 * Just make sure we have seen our old_write and synchronize
1532 * with any interrupts that come in.
1537 * If the tail page is still the same as what we think
1538 * it is, then it is up to us to update the tail
1541 if (tail_page == READ_ONCE(cpu_buffer->tail_page)) {
1542 /* Zero the write counter */
1543 unsigned long val = old_write & ~RB_WRITE_MASK;
1544 unsigned long eval = old_entries & ~RB_WRITE_MASK;
1547 * This will only succeed if an interrupt did
1548 * not come in and change it. In which case, we
1549 * do not want to modify it.
1551 * We add (void) to let the compiler know that we do not care
1552 * about the return value of these functions. We use the
1553 * cmpxchg to only update if an interrupt did not already
1554 * do it for us. If the cmpxchg fails, we don't care.
1556 (void)local_cmpxchg(&next_page->write, old_write, val);
1557 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
1560 * No need to worry about races with clearing out the commit.
1561 * it only can increment when a commit takes place. But that
1562 * only happens in the outer most nested commit.
1564 local_set(&next_page->page->commit, 0);
1566 /* Again, either we update tail_page or an interrupt does */
1567 (void)cmpxchg(&cpu_buffer->tail_page, tail_page, next_page);
1571 static void rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
1572 struct buffer_page *bpage)
1574 unsigned long val = (unsigned long)bpage;
1576 RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK);
1580 * rb_check_pages - integrity check of buffer pages
1581 * @cpu_buffer: CPU buffer with pages to test
1583 * As a safety measure we check to make sure the data pages have not
1586 static void rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
1588 struct list_head *head = rb_list_head(cpu_buffer->pages);
1589 struct list_head *tmp;
1591 if (RB_WARN_ON(cpu_buffer,
1592 rb_list_head(rb_list_head(head->next)->prev) != head))
1595 if (RB_WARN_ON(cpu_buffer,
1596 rb_list_head(rb_list_head(head->prev)->next) != head))
1599 for (tmp = rb_list_head(head->next); tmp != head; tmp = rb_list_head(tmp->next)) {
1600 if (RB_WARN_ON(cpu_buffer,
1601 rb_list_head(rb_list_head(tmp->next)->prev) != tmp))
1604 if (RB_WARN_ON(cpu_buffer,
1605 rb_list_head(rb_list_head(tmp->prev)->next) != tmp))
1610 static int __rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1611 long nr_pages, struct list_head *pages)
1613 struct buffer_page *bpage, *tmp;
1614 bool user_thread = current->mm != NULL;
1619 * Check if the available memory is there first.
1620 * Note, si_mem_available() only gives us a rough estimate of available
1621 * memory. It may not be accurate. But we don't care, we just want
1622 * to prevent doing any allocation when it is obvious that it is
1623 * not going to succeed.
1625 i = si_mem_available();
1630 * __GFP_RETRY_MAYFAIL flag makes sure that the allocation fails
1631 * gracefully without invoking oom-killer and the system is not
1634 mflags = GFP_KERNEL | __GFP_RETRY_MAYFAIL;
1637 * If a user thread allocates too much, and si_mem_available()
1638 * reports there's enough memory, even though there is not.
1639 * Make sure the OOM killer kills this thread. This can happen
1640 * even with RETRY_MAYFAIL because another task may be doing
1641 * an allocation after this task has taken all memory.
1642 * This is the task the OOM killer needs to take out during this
1643 * loop, even if it was triggered by an allocation somewhere else.
1646 set_current_oom_origin();
1647 for (i = 0; i < nr_pages; i++) {
1650 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1651 mflags, cpu_to_node(cpu_buffer->cpu));
1655 rb_check_bpage(cpu_buffer, bpage);
1657 list_add(&bpage->list, pages);
1659 page = alloc_pages_node(cpu_to_node(cpu_buffer->cpu), mflags, 0);
1662 bpage->page = page_address(page);
1663 rb_init_page(bpage->page);
1665 if (user_thread && fatal_signal_pending(current))
1669 clear_current_oom_origin();
1674 list_for_each_entry_safe(bpage, tmp, pages, list) {
1675 list_del_init(&bpage->list);
1676 free_buffer_page(bpage);
1679 clear_current_oom_origin();
1684 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1685 unsigned long nr_pages)
1691 if (__rb_allocate_pages(cpu_buffer, nr_pages, &pages))
1695 * The ring buffer page list is a circular list that does not
1696 * start and end with a list head. All page list items point to
1699 cpu_buffer->pages = pages.next;
1702 cpu_buffer->nr_pages = nr_pages;
1704 rb_check_pages(cpu_buffer);
1709 static struct ring_buffer_per_cpu *
1710 rb_allocate_cpu_buffer(struct trace_buffer *buffer, long nr_pages, int cpu)
1712 struct ring_buffer_per_cpu *cpu_buffer;
1713 struct buffer_page *bpage;
1717 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1718 GFP_KERNEL, cpu_to_node(cpu));
1722 cpu_buffer->cpu = cpu;
1723 cpu_buffer->buffer = buffer;
1724 raw_spin_lock_init(&cpu_buffer->reader_lock);
1725 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1726 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1727 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1728 init_completion(&cpu_buffer->update_done);
1729 init_irq_work(&cpu_buffer->irq_work.work, rb_wake_up_waiters);
1730 init_waitqueue_head(&cpu_buffer->irq_work.waiters);
1731 init_waitqueue_head(&cpu_buffer->irq_work.full_waiters);
1733 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1734 GFP_KERNEL, cpu_to_node(cpu));
1736 goto fail_free_buffer;
1738 rb_check_bpage(cpu_buffer, bpage);
1740 cpu_buffer->reader_page = bpage;
1741 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1743 goto fail_free_reader;
1744 bpage->page = page_address(page);
1745 rb_init_page(bpage->page);
1747 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1748 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1750 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1752 goto fail_free_reader;
1754 cpu_buffer->head_page
1755 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1756 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1758 rb_head_page_activate(cpu_buffer);
1763 free_buffer_page(cpu_buffer->reader_page);
1770 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1772 struct list_head *head = cpu_buffer->pages;
1773 struct buffer_page *bpage, *tmp;
1775 irq_work_sync(&cpu_buffer->irq_work.work);
1777 free_buffer_page(cpu_buffer->reader_page);
1780 rb_head_page_deactivate(cpu_buffer);
1782 list_for_each_entry_safe(bpage, tmp, head, list) {
1783 list_del_init(&bpage->list);
1784 free_buffer_page(bpage);
1786 bpage = list_entry(head, struct buffer_page, list);
1787 free_buffer_page(bpage);
1790 free_page((unsigned long)cpu_buffer->free_page);
1796 * __ring_buffer_alloc - allocate a new ring_buffer
1797 * @size: the size in bytes per cpu that is needed.
1798 * @flags: attributes to set for the ring buffer.
1799 * @key: ring buffer reader_lock_key.
1801 * Currently the only flag that is available is the RB_FL_OVERWRITE
1802 * flag. This flag means that the buffer will overwrite old data
1803 * when the buffer wraps. If this flag is not set, the buffer will
1804 * drop data when the tail hits the head.
1806 struct trace_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1807 struct lock_class_key *key)
1809 struct trace_buffer *buffer;
1815 /* keep it in its own cache line */
1816 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1821 if (!zalloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1822 goto fail_free_buffer;
1824 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1825 buffer->flags = flags;
1826 buffer->clock = trace_clock_local;
1827 buffer->reader_lock_key = key;
1829 init_irq_work(&buffer->irq_work.work, rb_wake_up_waiters);
1830 init_waitqueue_head(&buffer->irq_work.waiters);
1832 /* need at least two pages */
1836 buffer->cpus = nr_cpu_ids;
1838 bsize = sizeof(void *) * nr_cpu_ids;
1839 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1841 if (!buffer->buffers)
1842 goto fail_free_cpumask;
1844 cpu = raw_smp_processor_id();
1845 cpumask_set_cpu(cpu, buffer->cpumask);
1846 buffer->buffers[cpu] = rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1847 if (!buffer->buffers[cpu])
1848 goto fail_free_buffers;
1850 ret = cpuhp_state_add_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1852 goto fail_free_buffers;
1854 mutex_init(&buffer->mutex);
1859 for_each_buffer_cpu(buffer, cpu) {
1860 if (buffer->buffers[cpu])
1861 rb_free_cpu_buffer(buffer->buffers[cpu]);
1863 kfree(buffer->buffers);
1866 free_cpumask_var(buffer->cpumask);
1872 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1875 * ring_buffer_free - free a ring buffer.
1876 * @buffer: the buffer to free.
1879 ring_buffer_free(struct trace_buffer *buffer)
1883 cpuhp_state_remove_instance(CPUHP_TRACE_RB_PREPARE, &buffer->node);
1885 irq_work_sync(&buffer->irq_work.work);
1887 for_each_buffer_cpu(buffer, cpu)
1888 rb_free_cpu_buffer(buffer->buffers[cpu]);
1890 kfree(buffer->buffers);
1891 free_cpumask_var(buffer->cpumask);
1895 EXPORT_SYMBOL_GPL(ring_buffer_free);
1897 void ring_buffer_set_clock(struct trace_buffer *buffer,
1900 buffer->clock = clock;
1903 void ring_buffer_set_time_stamp_abs(struct trace_buffer *buffer, bool abs)
1905 buffer->time_stamp_abs = abs;
1908 bool ring_buffer_time_stamp_abs(struct trace_buffer *buffer)
1910 return buffer->time_stamp_abs;
1913 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1915 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1917 return local_read(&bpage->entries) & RB_WRITE_MASK;
1920 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1922 return local_read(&bpage->write) & RB_WRITE_MASK;
1926 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned long nr_pages)
1928 struct list_head *tail_page, *to_remove, *next_page;
1929 struct buffer_page *to_remove_page, *tmp_iter_page;
1930 struct buffer_page *last_page, *first_page;
1931 unsigned long nr_removed;
1932 unsigned long head_bit;
1937 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1938 atomic_inc(&cpu_buffer->record_disabled);
1940 * We don't race with the readers since we have acquired the reader
1941 * lock. We also don't race with writers after disabling recording.
1942 * This makes it easy to figure out the first and the last page to be
1943 * removed from the list. We unlink all the pages in between including
1944 * the first and last pages. This is done in a busy loop so that we
1945 * lose the least number of traces.
1946 * The pages are freed after we restart recording and unlock readers.
1948 tail_page = &cpu_buffer->tail_page->list;
1951 * tail page might be on reader page, we remove the next page
1952 * from the ring buffer
1954 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1955 tail_page = rb_list_head(tail_page->next);
1956 to_remove = tail_page;
1958 /* start of pages to remove */
1959 first_page = list_entry(rb_list_head(to_remove->next),
1960 struct buffer_page, list);
1962 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1963 to_remove = rb_list_head(to_remove)->next;
1964 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1966 /* Read iterators need to reset themselves when some pages removed */
1967 cpu_buffer->pages_removed += nr_removed;
1969 next_page = rb_list_head(to_remove)->next;
1972 * Now we remove all pages between tail_page and next_page.
1973 * Make sure that we have head_bit value preserved for the
1976 tail_page->next = (struct list_head *)((unsigned long)next_page |
1978 next_page = rb_list_head(next_page);
1979 next_page->prev = tail_page;
1981 /* make sure pages points to a valid page in the ring buffer */
1982 cpu_buffer->pages = next_page;
1984 /* update head page */
1986 cpu_buffer->head_page = list_entry(next_page,
1987 struct buffer_page, list);
1989 /* pages are removed, resume tracing and then free the pages */
1990 atomic_dec(&cpu_buffer->record_disabled);
1991 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1993 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1995 /* last buffer page to remove */
1996 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1998 tmp_iter_page = first_page;
2003 to_remove_page = tmp_iter_page;
2004 rb_inc_page(&tmp_iter_page);
2006 /* update the counters */
2007 page_entries = rb_page_entries(to_remove_page);
2010 * If something was added to this page, it was full
2011 * since it is not the tail page. So we deduct the
2012 * bytes consumed in ring buffer from here.
2013 * Increment overrun to account for the lost events.
2015 local_add(page_entries, &cpu_buffer->overrun);
2016 local_sub(rb_page_commit(to_remove_page), &cpu_buffer->entries_bytes);
2017 local_inc(&cpu_buffer->pages_lost);
2021 * We have already removed references to this list item, just
2022 * free up the buffer_page and its page
2024 free_buffer_page(to_remove_page);
2027 } while (to_remove_page != last_page);
2029 RB_WARN_ON(cpu_buffer, nr_removed);
2031 return nr_removed == 0;
2035 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
2037 struct list_head *pages = &cpu_buffer->new_pages;
2038 unsigned long flags;
2042 /* Can be called at early boot up, where interrupts must not been enabled */
2043 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2045 * We are holding the reader lock, so the reader page won't be swapped
2046 * in the ring buffer. Now we are racing with the writer trying to
2047 * move head page and the tail page.
2048 * We are going to adapt the reader page update process where:
2049 * 1. We first splice the start and end of list of new pages between
2050 * the head page and its previous page.
2051 * 2. We cmpxchg the prev_page->next to point from head page to the
2052 * start of new pages list.
2053 * 3. Finally, we update the head->prev to the end of new list.
2055 * We will try this process 10 times, to make sure that we don't keep
2061 struct list_head *head_page, *prev_page, *r;
2062 struct list_head *last_page, *first_page;
2063 struct list_head *head_page_with_bit;
2064 struct buffer_page *hpage = rb_set_head_page(cpu_buffer);
2068 head_page = &hpage->list;
2069 prev_page = head_page->prev;
2071 first_page = pages->next;
2072 last_page = pages->prev;
2074 head_page_with_bit = (struct list_head *)
2075 ((unsigned long)head_page | RB_PAGE_HEAD);
2077 last_page->next = head_page_with_bit;
2078 first_page->prev = prev_page;
2080 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
2082 if (r == head_page_with_bit) {
2084 * yay, we replaced the page pointer to our new list,
2085 * now, we just have to update to head page's prev
2086 * pointer to point to end of list
2088 head_page->prev = last_page;
2095 INIT_LIST_HEAD(pages);
2097 * If we weren't successful in adding in new pages, warn and stop
2100 RB_WARN_ON(cpu_buffer, !success);
2101 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2103 /* free pages if they weren't inserted */
2105 struct buffer_page *bpage, *tmp;
2106 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2108 list_del_init(&bpage->list);
2109 free_buffer_page(bpage);
2115 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
2119 if (cpu_buffer->nr_pages_to_update > 0)
2120 success = rb_insert_pages(cpu_buffer);
2122 success = rb_remove_pages(cpu_buffer,
2123 -cpu_buffer->nr_pages_to_update);
2126 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
2129 static void update_pages_handler(struct work_struct *work)
2131 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
2132 struct ring_buffer_per_cpu, update_pages_work);
2133 rb_update_pages(cpu_buffer);
2134 complete(&cpu_buffer->update_done);
2138 * ring_buffer_resize - resize the ring buffer
2139 * @buffer: the buffer to resize.
2140 * @size: the new size.
2141 * @cpu_id: the cpu buffer to resize
2143 * Minimum size is 2 * BUF_PAGE_SIZE.
2145 * Returns 0 on success and < 0 on failure.
2147 int ring_buffer_resize(struct trace_buffer *buffer, unsigned long size,
2150 struct ring_buffer_per_cpu *cpu_buffer;
2151 unsigned long nr_pages;
2155 * Always succeed at resizing a non-existent buffer:
2160 /* Make sure the requested buffer exists */
2161 if (cpu_id != RING_BUFFER_ALL_CPUS &&
2162 !cpumask_test_cpu(cpu_id, buffer->cpumask))
2165 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
2167 /* we need a minimum of two pages */
2171 /* prevent another thread from changing buffer sizes */
2172 mutex_lock(&buffer->mutex);
2173 atomic_inc(&buffer->resizing);
2175 if (cpu_id == RING_BUFFER_ALL_CPUS) {
2177 * Don't succeed if resizing is disabled, as a reader might be
2178 * manipulating the ring buffer and is expecting a sane state while
2181 for_each_buffer_cpu(buffer, cpu) {
2182 cpu_buffer = buffer->buffers[cpu];
2183 if (atomic_read(&cpu_buffer->resize_disabled)) {
2185 goto out_err_unlock;
2189 /* calculate the pages to update */
2190 for_each_buffer_cpu(buffer, cpu) {
2191 cpu_buffer = buffer->buffers[cpu];
2193 cpu_buffer->nr_pages_to_update = nr_pages -
2194 cpu_buffer->nr_pages;
2196 * nothing more to do for removing pages or no update
2198 if (cpu_buffer->nr_pages_to_update <= 0)
2201 * to add pages, make sure all new pages can be
2202 * allocated without receiving ENOMEM
2204 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2205 if (__rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2206 &cpu_buffer->new_pages)) {
2207 /* not enough memory for new pages */
2217 * Fire off all the required work handlers
2218 * We can't schedule on offline CPUs, but it's not necessary
2219 * since we can change their buffer sizes without any race.
2221 for_each_buffer_cpu(buffer, cpu) {
2222 cpu_buffer = buffer->buffers[cpu];
2223 if (!cpu_buffer->nr_pages_to_update)
2226 /* Can't run something on an offline CPU. */
2227 if (!cpu_online(cpu)) {
2228 rb_update_pages(cpu_buffer);
2229 cpu_buffer->nr_pages_to_update = 0;
2231 /* Run directly if possible. */
2233 if (cpu != smp_processor_id()) {
2235 schedule_work_on(cpu,
2236 &cpu_buffer->update_pages_work);
2238 update_pages_handler(&cpu_buffer->update_pages_work);
2244 /* wait for all the updates to complete */
2245 for_each_buffer_cpu(buffer, cpu) {
2246 cpu_buffer = buffer->buffers[cpu];
2247 if (!cpu_buffer->nr_pages_to_update)
2250 if (cpu_online(cpu))
2251 wait_for_completion(&cpu_buffer->update_done);
2252 cpu_buffer->nr_pages_to_update = 0;
2257 cpu_buffer = buffer->buffers[cpu_id];
2259 if (nr_pages == cpu_buffer->nr_pages)
2263 * Don't succeed if resizing is disabled, as a reader might be
2264 * manipulating the ring buffer and is expecting a sane state while
2267 if (atomic_read(&cpu_buffer->resize_disabled)) {
2269 goto out_err_unlock;
2272 cpu_buffer->nr_pages_to_update = nr_pages -
2273 cpu_buffer->nr_pages;
2275 INIT_LIST_HEAD(&cpu_buffer->new_pages);
2276 if (cpu_buffer->nr_pages_to_update > 0 &&
2277 __rb_allocate_pages(cpu_buffer, cpu_buffer->nr_pages_to_update,
2278 &cpu_buffer->new_pages)) {
2285 /* Can't run something on an offline CPU. */
2286 if (!cpu_online(cpu_id))
2287 rb_update_pages(cpu_buffer);
2289 /* Run directly if possible. */
2291 if (cpu_id == smp_processor_id()) {
2292 rb_update_pages(cpu_buffer);
2296 schedule_work_on(cpu_id,
2297 &cpu_buffer->update_pages_work);
2298 wait_for_completion(&cpu_buffer->update_done);
2302 cpu_buffer->nr_pages_to_update = 0;
2308 * The ring buffer resize can happen with the ring buffer
2309 * enabled, so that the update disturbs the tracing as little
2310 * as possible. But if the buffer is disabled, we do not need
2311 * to worry about that, and we can take the time to verify
2312 * that the buffer is not corrupt.
2314 if (atomic_read(&buffer->record_disabled)) {
2315 atomic_inc(&buffer->record_disabled);
2317 * Even though the buffer was disabled, we must make sure
2318 * that it is truly disabled before calling rb_check_pages.
2319 * There could have been a race between checking
2320 * record_disable and incrementing it.
2323 for_each_buffer_cpu(buffer, cpu) {
2324 cpu_buffer = buffer->buffers[cpu];
2325 rb_check_pages(cpu_buffer);
2327 atomic_dec(&buffer->record_disabled);
2330 atomic_dec(&buffer->resizing);
2331 mutex_unlock(&buffer->mutex);
2335 for_each_buffer_cpu(buffer, cpu) {
2336 struct buffer_page *bpage, *tmp;
2338 cpu_buffer = buffer->buffers[cpu];
2339 cpu_buffer->nr_pages_to_update = 0;
2341 if (list_empty(&cpu_buffer->new_pages))
2344 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
2346 list_del_init(&bpage->list);
2347 free_buffer_page(bpage);
2351 atomic_dec(&buffer->resizing);
2352 mutex_unlock(&buffer->mutex);
2355 EXPORT_SYMBOL_GPL(ring_buffer_resize);
2357 void ring_buffer_change_overwrite(struct trace_buffer *buffer, int val)
2359 mutex_lock(&buffer->mutex);
2361 buffer->flags |= RB_FL_OVERWRITE;
2363 buffer->flags &= ~RB_FL_OVERWRITE;
2364 mutex_unlock(&buffer->mutex);
2366 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
2368 static __always_inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
2370 return bpage->page->data + index;
2373 static __always_inline struct ring_buffer_event *
2374 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
2376 return __rb_page_index(cpu_buffer->reader_page,
2377 cpu_buffer->reader_page->read);
2380 static struct ring_buffer_event *
2381 rb_iter_head_event(struct ring_buffer_iter *iter)
2383 struct ring_buffer_event *event;
2384 struct buffer_page *iter_head_page = iter->head_page;
2385 unsigned long commit;
2388 if (iter->head != iter->next_event)
2392 * When the writer goes across pages, it issues a cmpxchg which
2393 * is a mb(), which will synchronize with the rmb here.
2394 * (see rb_tail_page_update() and __rb_reserve_next())
2396 commit = rb_page_commit(iter_head_page);
2399 /* An event needs to be at least 8 bytes in size */
2400 if (iter->head > commit - 8)
2403 event = __rb_page_index(iter_head_page, iter->head);
2404 length = rb_event_length(event);
2407 * READ_ONCE() doesn't work on functions and we don't want the
2408 * compiler doing any crazy optimizations with length.
2412 if ((iter->head + length) > commit || length > BUF_MAX_DATA_SIZE)
2413 /* Writer corrupted the read? */
2416 memcpy(iter->event, event, length);
2418 * If the page stamp is still the same after this rmb() then the
2419 * event was safely copied without the writer entering the page.
2423 /* Make sure the page didn't change since we read this */
2424 if (iter->page_stamp != iter_head_page->page->time_stamp ||
2425 commit > rb_page_commit(iter_head_page))
2428 iter->next_event = iter->head + length;
2431 /* Reset to the beginning */
2432 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2434 iter->next_event = 0;
2435 iter->missed_events = 1;
2439 /* Size is determined by what has been committed */
2440 static __always_inline unsigned rb_page_size(struct buffer_page *bpage)
2442 return rb_page_commit(bpage);
2445 static __always_inline unsigned
2446 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
2448 return rb_page_commit(cpu_buffer->commit_page);
2451 static __always_inline unsigned
2452 rb_event_index(struct ring_buffer_event *event)
2454 unsigned long addr = (unsigned long)event;
2456 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
2459 static void rb_inc_iter(struct ring_buffer_iter *iter)
2461 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2464 * The iterator could be on the reader page (it starts there).
2465 * But the head could have moved, since the reader was
2466 * found. Check for this case and assign the iterator
2467 * to the head page instead of next.
2469 if (iter->head_page == cpu_buffer->reader_page)
2470 iter->head_page = rb_set_head_page(cpu_buffer);
2472 rb_inc_page(&iter->head_page);
2474 iter->page_stamp = iter->read_stamp = iter->head_page->page->time_stamp;
2476 iter->next_event = 0;
2480 * rb_handle_head_page - writer hit the head page
2482 * Returns: +1 to retry page
2487 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
2488 struct buffer_page *tail_page,
2489 struct buffer_page *next_page)
2491 struct buffer_page *new_head;
2496 entries = rb_page_entries(next_page);
2499 * The hard part is here. We need to move the head
2500 * forward, and protect against both readers on
2501 * other CPUs and writers coming in via interrupts.
2503 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
2507 * type can be one of four:
2508 * NORMAL - an interrupt already moved it for us
2509 * HEAD - we are the first to get here.
2510 * UPDATE - we are the interrupt interrupting
2512 * MOVED - a reader on another CPU moved the next
2513 * pointer to its reader page. Give up
2520 * We changed the head to UPDATE, thus
2521 * it is our responsibility to update
2524 local_add(entries, &cpu_buffer->overrun);
2525 local_sub(rb_page_commit(next_page), &cpu_buffer->entries_bytes);
2526 local_inc(&cpu_buffer->pages_lost);
2529 * The entries will be zeroed out when we move the
2533 /* still more to do */
2536 case RB_PAGE_UPDATE:
2538 * This is an interrupt that interrupt the
2539 * previous update. Still more to do.
2542 case RB_PAGE_NORMAL:
2544 * An interrupt came in before the update
2545 * and processed this for us.
2546 * Nothing left to do.
2551 * The reader is on another CPU and just did
2552 * a swap with our next_page.
2557 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
2562 * Now that we are here, the old head pointer is
2563 * set to UPDATE. This will keep the reader from
2564 * swapping the head page with the reader page.
2565 * The reader (on another CPU) will spin till
2568 * We just need to protect against interrupts
2569 * doing the job. We will set the next pointer
2570 * to HEAD. After that, we set the old pointer
2571 * to NORMAL, but only if it was HEAD before.
2572 * otherwise we are an interrupt, and only
2573 * want the outer most commit to reset it.
2575 new_head = next_page;
2576 rb_inc_page(&new_head);
2578 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
2582 * Valid returns are:
2583 * HEAD - an interrupt came in and already set it.
2584 * NORMAL - One of two things:
2585 * 1) We really set it.
2586 * 2) A bunch of interrupts came in and moved
2587 * the page forward again.
2591 case RB_PAGE_NORMAL:
2595 RB_WARN_ON(cpu_buffer, 1);
2600 * It is possible that an interrupt came in,
2601 * set the head up, then more interrupts came in
2602 * and moved it again. When we get back here,
2603 * the page would have been set to NORMAL but we
2604 * just set it back to HEAD.
2606 * How do you detect this? Well, if that happened
2607 * the tail page would have moved.
2609 if (ret == RB_PAGE_NORMAL) {
2610 struct buffer_page *buffer_tail_page;
2612 buffer_tail_page = READ_ONCE(cpu_buffer->tail_page);
2614 * If the tail had moved passed next, then we need
2615 * to reset the pointer.
2617 if (buffer_tail_page != tail_page &&
2618 buffer_tail_page != next_page)
2619 rb_head_page_set_normal(cpu_buffer, new_head,
2625 * If this was the outer most commit (the one that
2626 * changed the original pointer from HEAD to UPDATE),
2627 * then it is up to us to reset it to NORMAL.
2629 if (type == RB_PAGE_HEAD) {
2630 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2633 if (RB_WARN_ON(cpu_buffer,
2634 ret != RB_PAGE_UPDATE))
2642 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2643 unsigned long tail, struct rb_event_info *info)
2645 struct buffer_page *tail_page = info->tail_page;
2646 struct ring_buffer_event *event;
2647 unsigned long length = info->length;
2650 * Only the event that crossed the page boundary
2651 * must fill the old tail_page with padding.
2653 if (tail >= BUF_PAGE_SIZE) {
2655 * If the page was filled, then we still need
2656 * to update the real_end. Reset it to zero
2657 * and the reader will ignore it.
2659 if (tail == BUF_PAGE_SIZE)
2660 tail_page->real_end = 0;
2662 local_sub(length, &tail_page->write);
2666 event = __rb_page_index(tail_page, tail);
2669 * Save the original length to the meta data.
2670 * This will be used by the reader to add lost event
2673 tail_page->real_end = tail;
2676 * If this event is bigger than the minimum size, then
2677 * we need to be careful that we don't subtract the
2678 * write counter enough to allow another writer to slip
2680 * We put in a discarded commit instead, to make sure
2681 * that this space is not used again, and this space will
2682 * not be accounted into 'entries_bytes'.
2684 * If we are less than the minimum size, we don't need to
2687 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2688 /* No room for any events */
2690 /* Mark the rest of the page with padding */
2691 rb_event_set_padding(event);
2693 /* Make sure the padding is visible before the write update */
2696 /* Set the write back to the previous setting */
2697 local_sub(length, &tail_page->write);
2701 /* Put in a discarded event */
2702 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2703 event->type_len = RINGBUF_TYPE_PADDING;
2704 /* time delta must be non zero */
2705 event->time_delta = 1;
2707 /* account for padding bytes */
2708 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2710 /* Make sure the padding is visible before the tail_page->write update */
2713 /* Set write to end of buffer */
2714 length = (tail + length) - BUF_PAGE_SIZE;
2715 local_sub(length, &tail_page->write);
2718 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer);
2721 * This is the slow path, force gcc not to inline it.
2723 static noinline struct ring_buffer_event *
2724 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2725 unsigned long tail, struct rb_event_info *info)
2727 struct buffer_page *tail_page = info->tail_page;
2728 struct buffer_page *commit_page = cpu_buffer->commit_page;
2729 struct trace_buffer *buffer = cpu_buffer->buffer;
2730 struct buffer_page *next_page;
2733 next_page = tail_page;
2735 rb_inc_page(&next_page);
2738 * If for some reason, we had an interrupt storm that made
2739 * it all the way around the buffer, bail, and warn
2742 if (unlikely(next_page == commit_page)) {
2743 local_inc(&cpu_buffer->commit_overrun);
2748 * This is where the fun begins!
2750 * We are fighting against races between a reader that
2751 * could be on another CPU trying to swap its reader
2752 * page with the buffer head.
2754 * We are also fighting against interrupts coming in and
2755 * moving the head or tail on us as well.
2757 * If the next page is the head page then we have filled
2758 * the buffer, unless the commit page is still on the
2761 if (rb_is_head_page(next_page, &tail_page->list)) {
2764 * If the commit is not on the reader page, then
2765 * move the header page.
2767 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2769 * If we are not in overwrite mode,
2770 * this is easy, just stop here.
2772 if (!(buffer->flags & RB_FL_OVERWRITE)) {
2773 local_inc(&cpu_buffer->dropped_events);
2777 ret = rb_handle_head_page(cpu_buffer,
2786 * We need to be careful here too. The
2787 * commit page could still be on the reader
2788 * page. We could have a small buffer, and
2789 * have filled up the buffer with events
2790 * from interrupts and such, and wrapped.
2792 * Note, if the tail page is also on the
2793 * reader_page, we let it move out.
2795 if (unlikely((cpu_buffer->commit_page !=
2796 cpu_buffer->tail_page) &&
2797 (cpu_buffer->commit_page ==
2798 cpu_buffer->reader_page))) {
2799 local_inc(&cpu_buffer->commit_overrun);
2805 rb_tail_page_update(cpu_buffer, tail_page, next_page);
2809 rb_reset_tail(cpu_buffer, tail, info);
2811 /* Commit what we have for now. */
2812 rb_end_commit(cpu_buffer);
2813 /* rb_end_commit() decs committing */
2814 local_inc(&cpu_buffer->committing);
2816 /* fail and let the caller try again */
2817 return ERR_PTR(-EAGAIN);
2821 rb_reset_tail(cpu_buffer, tail, info);
2827 static struct ring_buffer_event *
2828 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta, bool abs)
2831 event->type_len = RINGBUF_TYPE_TIME_STAMP;
2833 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
2835 /* Not the first event on the page, or not delta? */
2836 if (abs || rb_event_index(event)) {
2837 event->time_delta = delta & TS_MASK;
2838 event->array[0] = delta >> TS_SHIFT;
2840 /* nope, just zero it */
2841 event->time_delta = 0;
2842 event->array[0] = 0;
2845 return skip_time_extend(event);
2848 #ifndef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2849 static inline bool sched_clock_stable(void)
2856 rb_check_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2857 struct rb_event_info *info)
2861 WARN_ONCE(1, "Delta way too big! %llu ts=%llu before=%llu after=%llu write stamp=%llu\n%s",
2862 (unsigned long long)info->delta,
2863 (unsigned long long)info->ts,
2864 (unsigned long long)info->before,
2865 (unsigned long long)info->after,
2866 (unsigned long long)(rb_time_read(&cpu_buffer->write_stamp, &write_stamp) ? write_stamp : 0),
2867 sched_clock_stable() ? "" :
2868 "If you just came from a suspend/resume,\n"
2869 "please switch to the trace global clock:\n"
2870 " echo global > /sys/kernel/tracing/trace_clock\n"
2871 "or add trace_clock=global to the kernel command line\n");
2874 static void rb_add_timestamp(struct ring_buffer_per_cpu *cpu_buffer,
2875 struct ring_buffer_event **event,
2876 struct rb_event_info *info,
2878 unsigned int *length)
2880 bool abs = info->add_timestamp &
2881 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE);
2883 if (unlikely(info->delta > (1ULL << 59))) {
2885 * Some timers can use more than 59 bits, and when a timestamp
2886 * is added to the buffer, it will lose those bits.
2888 if (abs && (info->ts & TS_MSB)) {
2889 info->delta &= ABS_TS_MASK;
2891 /* did the clock go backwards */
2892 } else if (info->before == info->after && info->before > info->ts) {
2893 /* not interrupted */
2897 * This is possible with a recalibrating of the TSC.
2898 * Do not produce a call stack, but just report it.
2902 pr_warn("Ring buffer clock went backwards: %llu -> %llu\n",
2903 info->before, info->ts);
2906 rb_check_timestamp(cpu_buffer, info);
2910 *event = rb_add_time_stamp(*event, info->delta, abs);
2911 *length -= RB_LEN_TIME_EXTEND;
2916 * rb_update_event - update event type and data
2917 * @cpu_buffer: The per cpu buffer of the @event
2918 * @event: the event to update
2919 * @info: The info to update the @event with (contains length and delta)
2921 * Update the type and data fields of the @event. The length
2922 * is the actual size that is written to the ring buffer,
2923 * and with this, we can determine what to place into the
2927 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
2928 struct ring_buffer_event *event,
2929 struct rb_event_info *info)
2931 unsigned length = info->length;
2932 u64 delta = info->delta;
2933 unsigned int nest = local_read(&cpu_buffer->committing) - 1;
2935 if (!WARN_ON_ONCE(nest >= MAX_NEST))
2936 cpu_buffer->event_stamp[nest] = info->ts;
2939 * If we need to add a timestamp, then we
2940 * add it to the start of the reserved space.
2942 if (unlikely(info->add_timestamp))
2943 rb_add_timestamp(cpu_buffer, &event, info, &delta, &length);
2945 event->time_delta = delta;
2946 length -= RB_EVNT_HDR_SIZE;
2947 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
2948 event->type_len = 0;
2949 event->array[0] = length;
2951 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
2954 static unsigned rb_calculate_event_length(unsigned length)
2956 struct ring_buffer_event event; /* Used only for sizeof array */
2958 /* zero length can cause confusions */
2962 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2963 length += sizeof(event.array[0]);
2965 length += RB_EVNT_HDR_SIZE;
2966 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2969 * In case the time delta is larger than the 27 bits for it
2970 * in the header, we need to add a timestamp. If another
2971 * event comes in when trying to discard this one to increase
2972 * the length, then the timestamp will be added in the allocated
2973 * space of this event. If length is bigger than the size needed
2974 * for the TIME_EXTEND, then padding has to be used. The events
2975 * length must be either RB_LEN_TIME_EXTEND, or greater than or equal
2976 * to RB_LEN_TIME_EXTEND + 8, as 8 is the minimum size for padding.
2977 * As length is a multiple of 4, we only need to worry if it
2978 * is 12 (RB_LEN_TIME_EXTEND + 4).
2980 if (length == RB_LEN_TIME_EXTEND + RB_ALIGNMENT)
2981 length += RB_ALIGNMENT;
2986 static u64 rb_time_delta(struct ring_buffer_event *event)
2988 switch (event->type_len) {
2989 case RINGBUF_TYPE_PADDING:
2992 case RINGBUF_TYPE_TIME_EXTEND:
2993 return rb_event_time_stamp(event);
2995 case RINGBUF_TYPE_TIME_STAMP:
2998 case RINGBUF_TYPE_DATA:
2999 return event->time_delta;
3006 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
3007 struct ring_buffer_event *event)
3009 unsigned long new_index, old_index;
3010 struct buffer_page *bpage;
3015 new_index = rb_event_index(event);
3016 old_index = new_index + rb_event_ts_length(event);
3017 addr = (unsigned long)event;
3020 bpage = READ_ONCE(cpu_buffer->tail_page);
3022 delta = rb_time_delta(event);
3024 if (!rb_time_read(&cpu_buffer->write_stamp, &write_stamp))
3027 /* Make sure the write stamp is read before testing the location */
3030 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
3031 unsigned long write_mask =
3032 local_read(&bpage->write) & ~RB_WRITE_MASK;
3033 unsigned long event_length = rb_event_length(event);
3036 * For the before_stamp to be different than the write_stamp
3037 * to make sure that the next event adds an absolute
3038 * value and does not rely on the saved write stamp, which
3039 * is now going to be bogus.
3041 rb_time_set(&cpu_buffer->before_stamp, 0);
3043 /* Something came in, can't discard */
3044 if (!rb_time_cmpxchg(&cpu_buffer->write_stamp,
3045 write_stamp, write_stamp - delta))
3049 * If an event were to come in now, it would see that the
3050 * write_stamp and the before_stamp are different, and assume
3051 * that this event just added itself before updating
3052 * the write stamp. The interrupting event will fix the
3053 * write stamp for us, and use the before stamp as its delta.
3057 * This is on the tail page. It is possible that
3058 * a write could come in and move the tail page
3059 * and write to the next page. That is fine
3060 * because we just shorten what is on this page.
3062 old_index += write_mask;
3063 new_index += write_mask;
3065 /* caution: old_index gets updated on cmpxchg failure */
3066 if (local_try_cmpxchg(&bpage->write, &old_index, new_index)) {
3067 /* update counters */
3068 local_sub(event_length, &cpu_buffer->entries_bytes);
3073 /* could not discard */
3077 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
3079 local_inc(&cpu_buffer->committing);
3080 local_inc(&cpu_buffer->commits);
3083 static __always_inline void
3084 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
3086 unsigned long max_count;
3089 * We only race with interrupts and NMIs on this CPU.
3090 * If we own the commit event, then we can commit
3091 * all others that interrupted us, since the interruptions
3092 * are in stack format (they finish before they come
3093 * back to us). This allows us to do a simple loop to
3094 * assign the commit to the tail.
3097 max_count = cpu_buffer->nr_pages * 100;
3099 while (cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)) {
3100 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
3102 if (RB_WARN_ON(cpu_buffer,
3103 rb_is_reader_page(cpu_buffer->tail_page)))
3106 * No need for a memory barrier here, as the update
3107 * of the tail_page did it for this page.
3109 local_set(&cpu_buffer->commit_page->page->commit,
3110 rb_page_write(cpu_buffer->commit_page));
3111 rb_inc_page(&cpu_buffer->commit_page);
3112 /* add barrier to keep gcc from optimizing too much */
3115 while (rb_commit_index(cpu_buffer) !=
3116 rb_page_write(cpu_buffer->commit_page)) {
3118 /* Make sure the readers see the content of what is committed. */
3120 local_set(&cpu_buffer->commit_page->page->commit,
3121 rb_page_write(cpu_buffer->commit_page));
3122 RB_WARN_ON(cpu_buffer,
3123 local_read(&cpu_buffer->commit_page->page->commit) &
3128 /* again, keep gcc from optimizing */
3132 * If an interrupt came in just after the first while loop
3133 * and pushed the tail page forward, we will be left with
3134 * a dangling commit that will never go forward.
3136 if (unlikely(cpu_buffer->commit_page != READ_ONCE(cpu_buffer->tail_page)))
3140 static __always_inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
3142 unsigned long commits;
3144 if (RB_WARN_ON(cpu_buffer,
3145 !local_read(&cpu_buffer->committing)))
3149 commits = local_read(&cpu_buffer->commits);
3150 /* synchronize with interrupts */
3152 if (local_read(&cpu_buffer->committing) == 1)
3153 rb_set_commit_to_write(cpu_buffer);
3155 local_dec(&cpu_buffer->committing);
3157 /* synchronize with interrupts */
3161 * Need to account for interrupts coming in between the
3162 * updating of the commit page and the clearing of the
3163 * committing counter.
3165 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
3166 !local_read(&cpu_buffer->committing)) {
3167 local_inc(&cpu_buffer->committing);
3172 static inline void rb_event_discard(struct ring_buffer_event *event)
3174 if (extended_time(event))
3175 event = skip_time_extend(event);
3177 /* array[0] holds the actual length for the discarded event */
3178 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
3179 event->type_len = RINGBUF_TYPE_PADDING;
3180 /* time delta must be non zero */
3181 if (!event->time_delta)
3182 event->time_delta = 1;
3185 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer)
3187 local_inc(&cpu_buffer->entries);
3188 rb_end_commit(cpu_buffer);
3191 static __always_inline void
3192 rb_wakeups(struct trace_buffer *buffer, struct ring_buffer_per_cpu *cpu_buffer)
3194 if (buffer->irq_work.waiters_pending) {
3195 buffer->irq_work.waiters_pending = false;
3196 /* irq_work_queue() supplies it's own memory barriers */
3197 irq_work_queue(&buffer->irq_work.work);
3200 if (cpu_buffer->irq_work.waiters_pending) {
3201 cpu_buffer->irq_work.waiters_pending = false;
3202 /* irq_work_queue() supplies it's own memory barriers */
3203 irq_work_queue(&cpu_buffer->irq_work.work);
3206 if (cpu_buffer->last_pages_touch == local_read(&cpu_buffer->pages_touched))
3209 if (cpu_buffer->reader_page == cpu_buffer->commit_page)
3212 if (!cpu_buffer->irq_work.full_waiters_pending)
3215 cpu_buffer->last_pages_touch = local_read(&cpu_buffer->pages_touched);
3217 if (!full_hit(buffer, cpu_buffer->cpu, cpu_buffer->shortest_full))
3220 cpu_buffer->irq_work.wakeup_full = true;
3221 cpu_buffer->irq_work.full_waiters_pending = false;
3222 /* irq_work_queue() supplies it's own memory barriers */
3223 irq_work_queue(&cpu_buffer->irq_work.work);
3226 #ifdef CONFIG_RING_BUFFER_RECORD_RECURSION
3227 # define do_ring_buffer_record_recursion() \
3228 do_ftrace_record_recursion(_THIS_IP_, _RET_IP_)
3230 # define do_ring_buffer_record_recursion() do { } while (0)
3234 * The lock and unlock are done within a preempt disable section.
3235 * The current_context per_cpu variable can only be modified
3236 * by the current task between lock and unlock. But it can
3237 * be modified more than once via an interrupt. To pass this
3238 * information from the lock to the unlock without having to
3239 * access the 'in_interrupt()' functions again (which do show
3240 * a bit of overhead in something as critical as function tracing,
3241 * we use a bitmask trick.
3243 * bit 1 = NMI context
3244 * bit 2 = IRQ context
3245 * bit 3 = SoftIRQ context
3246 * bit 4 = normal context.
3248 * This works because this is the order of contexts that can
3249 * preempt other contexts. A SoftIRQ never preempts an IRQ
3252 * When the context is determined, the corresponding bit is
3253 * checked and set (if it was set, then a recursion of that context
3256 * On unlock, we need to clear this bit. To do so, just subtract
3257 * 1 from the current_context and AND it to itself.
3261 * 101 & 100 = 100 (clearing bit zero)
3264 * 1010 & 1001 = 1000 (clearing bit 1)
3266 * The least significant bit can be cleared this way, and it
3267 * just so happens that it is the same bit corresponding to
3268 * the current context.
3270 * Now the TRANSITION bit breaks the above slightly. The TRANSITION bit
3271 * is set when a recursion is detected at the current context, and if
3272 * the TRANSITION bit is already set, it will fail the recursion.
3273 * This is needed because there's a lag between the changing of
3274 * interrupt context and updating the preempt count. In this case,
3275 * a false positive will be found. To handle this, one extra recursion
3276 * is allowed, and this is done by the TRANSITION bit. If the TRANSITION
3277 * bit is already set, then it is considered a recursion and the function
3278 * ends. Otherwise, the TRANSITION bit is set, and that bit is returned.
3280 * On the trace_recursive_unlock(), the TRANSITION bit will be the first
3281 * to be cleared. Even if it wasn't the context that set it. That is,
3282 * if an interrupt comes in while NORMAL bit is set and the ring buffer
3283 * is called before preempt_count() is updated, since the check will
3284 * be on the NORMAL bit, the TRANSITION bit will then be set. If an
3285 * NMI then comes in, it will set the NMI bit, but when the NMI code
3286 * does the trace_recursive_unlock() it will clear the TRANSITION bit
3287 * and leave the NMI bit set. But this is fine, because the interrupt
3288 * code that set the TRANSITION bit will then clear the NMI bit when it
3289 * calls trace_recursive_unlock(). If another NMI comes in, it will
3290 * set the TRANSITION bit and continue.
3292 * Note: The TRANSITION bit only handles a single transition between context.
3295 static __always_inline bool
3296 trace_recursive_lock(struct ring_buffer_per_cpu *cpu_buffer)
3298 unsigned int val = cpu_buffer->current_context;
3299 int bit = interrupt_context_level();
3301 bit = RB_CTX_NORMAL - bit;
3303 if (unlikely(val & (1 << (bit + cpu_buffer->nest)))) {
3305 * It is possible that this was called by transitioning
3306 * between interrupt context, and preempt_count() has not
3307 * been updated yet. In this case, use the TRANSITION bit.
3309 bit = RB_CTX_TRANSITION;
3310 if (val & (1 << (bit + cpu_buffer->nest))) {
3311 do_ring_buffer_record_recursion();
3316 val |= (1 << (bit + cpu_buffer->nest));
3317 cpu_buffer->current_context = val;
3322 static __always_inline void
3323 trace_recursive_unlock(struct ring_buffer_per_cpu *cpu_buffer)
3325 cpu_buffer->current_context &=
3326 cpu_buffer->current_context - (1 << cpu_buffer->nest);
3329 /* The recursive locking above uses 5 bits */
3330 #define NESTED_BITS 5
3333 * ring_buffer_nest_start - Allow to trace while nested
3334 * @buffer: The ring buffer to modify
3336 * The ring buffer has a safety mechanism to prevent recursion.
3337 * But there may be a case where a trace needs to be done while
3338 * tracing something else. In this case, calling this function
3339 * will allow this function to nest within a currently active
3340 * ring_buffer_lock_reserve().
3342 * Call this function before calling another ring_buffer_lock_reserve() and
3343 * call ring_buffer_nest_end() after the nested ring_buffer_unlock_commit().
3345 void ring_buffer_nest_start(struct trace_buffer *buffer)
3347 struct ring_buffer_per_cpu *cpu_buffer;
3350 /* Enabled by ring_buffer_nest_end() */
3351 preempt_disable_notrace();
3352 cpu = raw_smp_processor_id();
3353 cpu_buffer = buffer->buffers[cpu];
3354 /* This is the shift value for the above recursive locking */
3355 cpu_buffer->nest += NESTED_BITS;
3359 * ring_buffer_nest_end - Allow to trace while nested
3360 * @buffer: The ring buffer to modify
3362 * Must be called after ring_buffer_nest_start() and after the
3363 * ring_buffer_unlock_commit().
3365 void ring_buffer_nest_end(struct trace_buffer *buffer)
3367 struct ring_buffer_per_cpu *cpu_buffer;
3370 /* disabled by ring_buffer_nest_start() */
3371 cpu = raw_smp_processor_id();
3372 cpu_buffer = buffer->buffers[cpu];
3373 /* This is the shift value for the above recursive locking */
3374 cpu_buffer->nest -= NESTED_BITS;
3375 preempt_enable_notrace();
3379 * ring_buffer_unlock_commit - commit a reserved
3380 * @buffer: The buffer to commit to
3382 * This commits the data to the ring buffer, and releases any locks held.
3384 * Must be paired with ring_buffer_lock_reserve.
3386 int ring_buffer_unlock_commit(struct trace_buffer *buffer)
3388 struct ring_buffer_per_cpu *cpu_buffer;
3389 int cpu = raw_smp_processor_id();
3391 cpu_buffer = buffer->buffers[cpu];
3393 rb_commit(cpu_buffer);
3395 rb_wakeups(buffer, cpu_buffer);
3397 trace_recursive_unlock(cpu_buffer);
3399 preempt_enable_notrace();
3403 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
3405 /* Special value to validate all deltas on a page. */
3406 #define CHECK_FULL_PAGE 1L
3408 #ifdef CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS
3409 static void dump_buffer_page(struct buffer_data_page *bpage,
3410 struct rb_event_info *info,
3413 struct ring_buffer_event *event;
3417 ts = bpage->time_stamp;
3418 pr_warn(" [%lld] PAGE TIME STAMP\n", ts);
3420 for (e = 0; e < tail; e += rb_event_length(event)) {
3422 event = (struct ring_buffer_event *)(bpage->data + e);
3424 switch (event->type_len) {
3426 case RINGBUF_TYPE_TIME_EXTEND:
3427 delta = rb_event_time_stamp(event);
3429 pr_warn(" [%lld] delta:%lld TIME EXTEND\n", ts, delta);
3432 case RINGBUF_TYPE_TIME_STAMP:
3433 delta = rb_event_time_stamp(event);
3434 ts = rb_fix_abs_ts(delta, ts);
3435 pr_warn(" [%lld] absolute:%lld TIME STAMP\n", ts, delta);
3438 case RINGBUF_TYPE_PADDING:
3439 ts += event->time_delta;
3440 pr_warn(" [%lld] delta:%d PADDING\n", ts, event->time_delta);
3443 case RINGBUF_TYPE_DATA:
3444 ts += event->time_delta;
3445 pr_warn(" [%lld] delta:%d\n", ts, event->time_delta);
3454 static DEFINE_PER_CPU(atomic_t, checking);
3455 static atomic_t ts_dump;
3458 * Check if the current event time stamp matches the deltas on
3461 static void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3462 struct rb_event_info *info,
3465 struct ring_buffer_event *event;
3466 struct buffer_data_page *bpage;
3471 bpage = info->tail_page->page;
3473 if (tail == CHECK_FULL_PAGE) {
3475 tail = local_read(&bpage->commit);
3476 } else if (info->add_timestamp &
3477 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE)) {
3478 /* Ignore events with absolute time stamps */
3483 * Do not check the first event (skip possible extends too).
3484 * Also do not check if previous events have not been committed.
3486 if (tail <= 8 || tail > local_read(&bpage->commit))
3490 * If this interrupted another event,
3492 if (atomic_inc_return(this_cpu_ptr(&checking)) != 1)
3495 ts = bpage->time_stamp;
3497 for (e = 0; e < tail; e += rb_event_length(event)) {
3499 event = (struct ring_buffer_event *)(bpage->data + e);
3501 switch (event->type_len) {
3503 case RINGBUF_TYPE_TIME_EXTEND:
3504 delta = rb_event_time_stamp(event);
3508 case RINGBUF_TYPE_TIME_STAMP:
3509 delta = rb_event_time_stamp(event);
3510 ts = rb_fix_abs_ts(delta, ts);
3513 case RINGBUF_TYPE_PADDING:
3514 if (event->time_delta == 1)
3517 case RINGBUF_TYPE_DATA:
3518 ts += event->time_delta;
3522 RB_WARN_ON(cpu_buffer, 1);
3525 if ((full && ts > info->ts) ||
3526 (!full && ts + info->delta != info->ts)) {
3527 /* If another report is happening, ignore this one */
3528 if (atomic_inc_return(&ts_dump) != 1) {
3529 atomic_dec(&ts_dump);
3532 atomic_inc(&cpu_buffer->record_disabled);
3533 /* There's some cases in boot up that this can happen */
3534 WARN_ON_ONCE(system_state != SYSTEM_BOOTING);
3535 pr_warn("[CPU: %d]TIME DOES NOT MATCH expected:%lld actual:%lld delta:%lld before:%lld after:%lld%s\n",
3537 ts + info->delta, info->ts, info->delta,
3538 info->before, info->after,
3539 full ? " (full)" : "");
3540 dump_buffer_page(bpage, info, tail);
3541 atomic_dec(&ts_dump);
3542 /* Do not re-enable checking */
3546 atomic_dec(this_cpu_ptr(&checking));
3549 static inline void check_buffer(struct ring_buffer_per_cpu *cpu_buffer,
3550 struct rb_event_info *info,
3554 #endif /* CONFIG_RING_BUFFER_VALIDATE_TIME_DELTAS */
3556 static struct ring_buffer_event *
3557 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
3558 struct rb_event_info *info)
3560 struct ring_buffer_event *event;
3561 struct buffer_page *tail_page;
3562 unsigned long tail, write, w;
3566 /* Don't let the compiler play games with cpu_buffer->tail_page */
3567 tail_page = info->tail_page = READ_ONCE(cpu_buffer->tail_page);
3569 /*A*/ w = local_read(&tail_page->write) & RB_WRITE_MASK;
3571 b_ok = rb_time_read(&cpu_buffer->before_stamp, &info->before);
3572 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3574 info->ts = rb_time_stamp(cpu_buffer->buffer);
3576 if ((info->add_timestamp & RB_ADD_STAMP_ABSOLUTE)) {
3577 info->delta = info->ts;
3580 * If interrupting an event time update, we may need an
3581 * absolute timestamp.
3582 * Don't bother if this is the start of a new page (w == 0).
3584 if (unlikely(!a_ok || !b_ok || (info->before != info->after && w))) {
3585 info->add_timestamp |= RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND;
3586 info->length += RB_LEN_TIME_EXTEND;
3588 info->delta = info->ts - info->after;
3589 if (unlikely(test_time_stamp(info->delta))) {
3590 info->add_timestamp |= RB_ADD_STAMP_EXTEND;
3591 info->length += RB_LEN_TIME_EXTEND;
3596 /*B*/ rb_time_set(&cpu_buffer->before_stamp, info->ts);
3598 /*C*/ write = local_add_return(info->length, &tail_page->write);
3600 /* set write to only the index of the write */
3601 write &= RB_WRITE_MASK;
3603 tail = write - info->length;
3605 /* See if we shot pass the end of this buffer page */
3606 if (unlikely(write > BUF_PAGE_SIZE)) {
3607 check_buffer(cpu_buffer, info, CHECK_FULL_PAGE);
3608 return rb_move_tail(cpu_buffer, tail, info);
3611 if (likely(tail == w)) {
3615 /* Nothing interrupted us between A and C */
3616 /*D*/ rb_time_set(&cpu_buffer->write_stamp, info->ts);
3618 /*E*/ s_ok = rb_time_read(&cpu_buffer->before_stamp, &save_before);
3619 RB_WARN_ON(cpu_buffer, !s_ok);
3620 if (likely(!(info->add_timestamp &
3621 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3622 /* This did not interrupt any time update */
3623 info->delta = info->ts - info->after;
3625 /* Just use full timestamp for interrupting event */
3626 info->delta = info->ts;
3628 check_buffer(cpu_buffer, info, tail);
3629 if (unlikely(info->ts != save_before)) {
3630 /* SLOW PATH - Interrupted between C and E */
3632 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3633 RB_WARN_ON(cpu_buffer, !a_ok);
3635 /* Write stamp must only go forward */
3636 if (save_before > info->after) {
3638 * We do not care about the result, only that
3639 * it gets updated atomically.
3641 (void)rb_time_cmpxchg(&cpu_buffer->write_stamp,
3642 info->after, save_before);
3647 /* SLOW PATH - Interrupted between A and C */
3648 a_ok = rb_time_read(&cpu_buffer->write_stamp, &info->after);
3649 /* Was interrupted before here, write_stamp must be valid */
3650 RB_WARN_ON(cpu_buffer, !a_ok);
3651 ts = rb_time_stamp(cpu_buffer->buffer);
3653 /*E*/ if (write == (local_read(&tail_page->write) & RB_WRITE_MASK) &&
3655 rb_time_cmpxchg(&cpu_buffer->write_stamp,
3657 /* Nothing came after this event between C and E */
3658 info->delta = ts - info->after;
3661 * Interrupted between C and E:
3662 * Lost the previous events time stamp. Just set the
3663 * delta to zero, and this will be the same time as
3664 * the event this event interrupted. And the events that
3665 * came after this will still be correct (as they would
3666 * have built their delta on the previous event.
3671 info->add_timestamp &= ~RB_ADD_STAMP_FORCE;
3675 * If this is the first commit on the page, then it has the same
3676 * timestamp as the page itself.
3678 if (unlikely(!tail && !(info->add_timestamp &
3679 (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_ABSOLUTE))))
3682 /* We reserved something on the buffer */
3684 event = __rb_page_index(tail_page, tail);
3685 rb_update_event(cpu_buffer, event, info);
3687 local_inc(&tail_page->entries);
3690 * If this is the first commit on the page, then update
3693 if (unlikely(!tail))
3694 tail_page->page->time_stamp = info->ts;
3696 /* account for these added bytes */
3697 local_add(info->length, &cpu_buffer->entries_bytes);
3702 static __always_inline struct ring_buffer_event *
3703 rb_reserve_next_event(struct trace_buffer *buffer,
3704 struct ring_buffer_per_cpu *cpu_buffer,
3705 unsigned long length)
3707 struct ring_buffer_event *event;
3708 struct rb_event_info info;
3712 rb_start_commit(cpu_buffer);
3713 /* The commit page can not change after this */
3715 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3717 * Due to the ability to swap a cpu buffer from a buffer
3718 * it is possible it was swapped before we committed.
3719 * (committing stops a swap). We check for it here and
3720 * if it happened, we have to fail the write.
3723 if (unlikely(READ_ONCE(cpu_buffer->buffer) != buffer)) {
3724 local_dec(&cpu_buffer->committing);
3725 local_dec(&cpu_buffer->commits);
3730 info.length = rb_calculate_event_length(length);
3732 if (ring_buffer_time_stamp_abs(cpu_buffer->buffer)) {
3733 add_ts_default = RB_ADD_STAMP_ABSOLUTE;
3734 info.length += RB_LEN_TIME_EXTEND;
3736 add_ts_default = RB_ADD_STAMP_NONE;
3740 info.add_timestamp = add_ts_default;
3744 * We allow for interrupts to reenter here and do a trace.
3745 * If one does, it will cause this original code to loop
3746 * back here. Even with heavy interrupts happening, this
3747 * should only happen a few times in a row. If this happens
3748 * 1000 times in a row, there must be either an interrupt
3749 * storm or we have something buggy.
3752 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
3755 event = __rb_reserve_next(cpu_buffer, &info);
3757 if (unlikely(PTR_ERR(event) == -EAGAIN)) {
3758 if (info.add_timestamp & (RB_ADD_STAMP_FORCE | RB_ADD_STAMP_EXTEND))
3759 info.length -= RB_LEN_TIME_EXTEND;
3766 rb_end_commit(cpu_buffer);
3771 * ring_buffer_lock_reserve - reserve a part of the buffer
3772 * @buffer: the ring buffer to reserve from
3773 * @length: the length of the data to reserve (excluding event header)
3775 * Returns a reserved event on the ring buffer to copy directly to.
3776 * The user of this interface will need to get the body to write into
3777 * and can use the ring_buffer_event_data() interface.
3779 * The length is the length of the data needed, not the event length
3780 * which also includes the event header.
3782 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
3783 * If NULL is returned, then nothing has been allocated or locked.
3785 struct ring_buffer_event *
3786 ring_buffer_lock_reserve(struct trace_buffer *buffer, unsigned long length)
3788 struct ring_buffer_per_cpu *cpu_buffer;
3789 struct ring_buffer_event *event;
3792 /* If we are tracing schedule, we don't want to recurse */
3793 preempt_disable_notrace();
3795 if (unlikely(atomic_read(&buffer->record_disabled)))
3798 cpu = raw_smp_processor_id();
3800 if (unlikely(!cpumask_test_cpu(cpu, buffer->cpumask)))
3803 cpu_buffer = buffer->buffers[cpu];
3805 if (unlikely(atomic_read(&cpu_buffer->record_disabled)))
3808 if (unlikely(length > BUF_MAX_DATA_SIZE))
3811 if (unlikely(trace_recursive_lock(cpu_buffer)))
3814 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3821 trace_recursive_unlock(cpu_buffer);
3823 preempt_enable_notrace();
3826 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
3829 * Decrement the entries to the page that an event is on.
3830 * The event does not even need to exist, only the pointer
3831 * to the page it is on. This may only be called before the commit
3835 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
3836 struct ring_buffer_event *event)
3838 unsigned long addr = (unsigned long)event;
3839 struct buffer_page *bpage = cpu_buffer->commit_page;
3840 struct buffer_page *start;
3844 /* Do the likely case first */
3845 if (likely(bpage->page == (void *)addr)) {
3846 local_dec(&bpage->entries);
3851 * Because the commit page may be on the reader page we
3852 * start with the next page and check the end loop there.
3854 rb_inc_page(&bpage);
3857 if (bpage->page == (void *)addr) {
3858 local_dec(&bpage->entries);
3861 rb_inc_page(&bpage);
3862 } while (bpage != start);
3864 /* commit not part of this buffer?? */
3865 RB_WARN_ON(cpu_buffer, 1);
3869 * ring_buffer_discard_commit - discard an event that has not been committed
3870 * @buffer: the ring buffer
3871 * @event: non committed event to discard
3873 * Sometimes an event that is in the ring buffer needs to be ignored.
3874 * This function lets the user discard an event in the ring buffer
3875 * and then that event will not be read later.
3877 * This function only works if it is called before the item has been
3878 * committed. It will try to free the event from the ring buffer
3879 * if another event has not been added behind it.
3881 * If another event has been added behind it, it will set the event
3882 * up as discarded, and perform the commit.
3884 * If this function is called, do not call ring_buffer_unlock_commit on
3887 void ring_buffer_discard_commit(struct trace_buffer *buffer,
3888 struct ring_buffer_event *event)
3890 struct ring_buffer_per_cpu *cpu_buffer;
3893 /* The event is discarded regardless */
3894 rb_event_discard(event);
3896 cpu = smp_processor_id();
3897 cpu_buffer = buffer->buffers[cpu];
3900 * This must only be called if the event has not been
3901 * committed yet. Thus we can assume that preemption
3902 * is still disabled.
3904 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
3906 rb_decrement_entry(cpu_buffer, event);
3907 if (rb_try_to_discard(cpu_buffer, event))
3911 rb_end_commit(cpu_buffer);
3913 trace_recursive_unlock(cpu_buffer);
3915 preempt_enable_notrace();
3918 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
3921 * ring_buffer_write - write data to the buffer without reserving
3922 * @buffer: The ring buffer to write to.
3923 * @length: The length of the data being written (excluding the event header)
3924 * @data: The data to write to the buffer.
3926 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
3927 * one function. If you already have the data to write to the buffer, it
3928 * may be easier to simply call this function.
3930 * Note, like ring_buffer_lock_reserve, the length is the length of the data
3931 * and not the length of the event which would hold the header.
3933 int ring_buffer_write(struct trace_buffer *buffer,
3934 unsigned long length,
3937 struct ring_buffer_per_cpu *cpu_buffer;
3938 struct ring_buffer_event *event;
3943 preempt_disable_notrace();
3945 if (atomic_read(&buffer->record_disabled))
3948 cpu = raw_smp_processor_id();
3950 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3953 cpu_buffer = buffer->buffers[cpu];
3955 if (atomic_read(&cpu_buffer->record_disabled))
3958 if (length > BUF_MAX_DATA_SIZE)
3961 if (unlikely(trace_recursive_lock(cpu_buffer)))
3964 event = rb_reserve_next_event(buffer, cpu_buffer, length);
3968 body = rb_event_data(event);
3970 memcpy(body, data, length);
3972 rb_commit(cpu_buffer);
3974 rb_wakeups(buffer, cpu_buffer);
3979 trace_recursive_unlock(cpu_buffer);
3982 preempt_enable_notrace();
3986 EXPORT_SYMBOL_GPL(ring_buffer_write);
3988 static bool rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
3990 struct buffer_page *reader = cpu_buffer->reader_page;
3991 struct buffer_page *head = rb_set_head_page(cpu_buffer);
3992 struct buffer_page *commit = cpu_buffer->commit_page;
3994 /* In case of error, head will be NULL */
3995 if (unlikely(!head))
3998 /* Reader should exhaust content in reader page */
3999 if (reader->read != rb_page_commit(reader))
4003 * If writers are committing on the reader page, knowing all
4004 * committed content has been read, the ring buffer is empty.
4006 if (commit == reader)
4010 * If writers are committing on a page other than reader page
4011 * and head page, there should always be content to read.
4017 * Writers are committing on the head page, we just need
4018 * to care about there're committed data, and the reader will
4019 * swap reader page with head page when it is to read data.
4021 return rb_page_commit(commit) == 0;
4025 * ring_buffer_record_disable - stop all writes into the buffer
4026 * @buffer: The ring buffer to stop writes to.
4028 * This prevents all writes to the buffer. Any attempt to write
4029 * to the buffer after this will fail and return NULL.
4031 * The caller should call synchronize_rcu() after this.
4033 void ring_buffer_record_disable(struct trace_buffer *buffer)
4035 atomic_inc(&buffer->record_disabled);
4037 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
4040 * ring_buffer_record_enable - enable writes to the buffer
4041 * @buffer: The ring buffer to enable writes
4043 * Note, multiple disables will need the same number of enables
4044 * to truly enable the writing (much like preempt_disable).
4046 void ring_buffer_record_enable(struct trace_buffer *buffer)
4048 atomic_dec(&buffer->record_disabled);
4050 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
4053 * ring_buffer_record_off - stop all writes into the buffer
4054 * @buffer: The ring buffer to stop writes to.
4056 * This prevents all writes to the buffer. Any attempt to write
4057 * to the buffer after this will fail and return NULL.
4059 * This is different than ring_buffer_record_disable() as
4060 * it works like an on/off switch, where as the disable() version
4061 * must be paired with a enable().
4063 void ring_buffer_record_off(struct trace_buffer *buffer)
4066 unsigned int new_rd;
4068 rd = atomic_read(&buffer->record_disabled);
4070 new_rd = rd | RB_BUFFER_OFF;
4071 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4073 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
4076 * ring_buffer_record_on - restart writes into the buffer
4077 * @buffer: The ring buffer to start writes to.
4079 * This enables all writes to the buffer that was disabled by
4080 * ring_buffer_record_off().
4082 * This is different than ring_buffer_record_enable() as
4083 * it works like an on/off switch, where as the enable() version
4084 * must be paired with a disable().
4086 void ring_buffer_record_on(struct trace_buffer *buffer)
4089 unsigned int new_rd;
4091 rd = atomic_read(&buffer->record_disabled);
4093 new_rd = rd & ~RB_BUFFER_OFF;
4094 } while (!atomic_try_cmpxchg(&buffer->record_disabled, &rd, new_rd));
4096 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
4099 * ring_buffer_record_is_on - return true if the ring buffer can write
4100 * @buffer: The ring buffer to see if write is enabled
4102 * Returns true if the ring buffer is in a state that it accepts writes.
4104 bool ring_buffer_record_is_on(struct trace_buffer *buffer)
4106 return !atomic_read(&buffer->record_disabled);
4110 * ring_buffer_record_is_set_on - return true if the ring buffer is set writable
4111 * @buffer: The ring buffer to see if write is set enabled
4113 * Returns true if the ring buffer is set writable by ring_buffer_record_on().
4114 * Note that this does NOT mean it is in a writable state.
4116 * It may return true when the ring buffer has been disabled by
4117 * ring_buffer_record_disable(), as that is a temporary disabling of
4120 bool ring_buffer_record_is_set_on(struct trace_buffer *buffer)
4122 return !(atomic_read(&buffer->record_disabled) & RB_BUFFER_OFF);
4126 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
4127 * @buffer: The ring buffer to stop writes to.
4128 * @cpu: The CPU buffer to stop
4130 * This prevents all writes to the buffer. Any attempt to write
4131 * to the buffer after this will fail and return NULL.
4133 * The caller should call synchronize_rcu() after this.
4135 void ring_buffer_record_disable_cpu(struct trace_buffer *buffer, int cpu)
4137 struct ring_buffer_per_cpu *cpu_buffer;
4139 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4142 cpu_buffer = buffer->buffers[cpu];
4143 atomic_inc(&cpu_buffer->record_disabled);
4145 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
4148 * ring_buffer_record_enable_cpu - enable writes to the buffer
4149 * @buffer: The ring buffer to enable writes
4150 * @cpu: The CPU to enable.
4152 * Note, multiple disables will need the same number of enables
4153 * to truly enable the writing (much like preempt_disable).
4155 void ring_buffer_record_enable_cpu(struct trace_buffer *buffer, int cpu)
4157 struct ring_buffer_per_cpu *cpu_buffer;
4159 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4162 cpu_buffer = buffer->buffers[cpu];
4163 atomic_dec(&cpu_buffer->record_disabled);
4165 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
4168 * The total entries in the ring buffer is the running counter
4169 * of entries entered into the ring buffer, minus the sum of
4170 * the entries read from the ring buffer and the number of
4171 * entries that were overwritten.
4173 static inline unsigned long
4174 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
4176 return local_read(&cpu_buffer->entries) -
4177 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
4181 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
4182 * @buffer: The ring buffer
4183 * @cpu: The per CPU buffer to read from.
4185 u64 ring_buffer_oldest_event_ts(struct trace_buffer *buffer, int cpu)
4187 unsigned long flags;
4188 struct ring_buffer_per_cpu *cpu_buffer;
4189 struct buffer_page *bpage;
4192 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4195 cpu_buffer = buffer->buffers[cpu];
4196 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4198 * if the tail is on reader_page, oldest time stamp is on the reader
4201 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
4202 bpage = cpu_buffer->reader_page;
4204 bpage = rb_set_head_page(cpu_buffer);
4206 ret = bpage->page->time_stamp;
4207 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4211 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
4214 * ring_buffer_bytes_cpu - get the number of bytes unconsumed in a cpu buffer
4215 * @buffer: The ring buffer
4216 * @cpu: The per CPU buffer to read from.
4218 unsigned long ring_buffer_bytes_cpu(struct trace_buffer *buffer, int cpu)
4220 struct ring_buffer_per_cpu *cpu_buffer;
4223 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4226 cpu_buffer = buffer->buffers[cpu];
4227 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
4231 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
4234 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
4235 * @buffer: The ring buffer
4236 * @cpu: The per CPU buffer to get the entries from.
4238 unsigned long ring_buffer_entries_cpu(struct trace_buffer *buffer, int cpu)
4240 struct ring_buffer_per_cpu *cpu_buffer;
4242 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4245 cpu_buffer = buffer->buffers[cpu];
4247 return rb_num_of_entries(cpu_buffer);
4249 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
4252 * ring_buffer_overrun_cpu - get the number of overruns caused by the ring
4253 * buffer wrapping around (only if RB_FL_OVERWRITE is on).
4254 * @buffer: The ring buffer
4255 * @cpu: The per CPU buffer to get the number of overruns from
4257 unsigned long ring_buffer_overrun_cpu(struct trace_buffer *buffer, int cpu)
4259 struct ring_buffer_per_cpu *cpu_buffer;
4262 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4265 cpu_buffer = buffer->buffers[cpu];
4266 ret = local_read(&cpu_buffer->overrun);
4270 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
4273 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by
4274 * commits failing due to the buffer wrapping around while there are uncommitted
4275 * events, such as during an interrupt storm.
4276 * @buffer: The ring buffer
4277 * @cpu: The per CPU buffer to get the number of overruns from
4280 ring_buffer_commit_overrun_cpu(struct trace_buffer *buffer, int cpu)
4282 struct ring_buffer_per_cpu *cpu_buffer;
4285 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4288 cpu_buffer = buffer->buffers[cpu];
4289 ret = local_read(&cpu_buffer->commit_overrun);
4293 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
4296 * ring_buffer_dropped_events_cpu - get the number of dropped events caused by
4297 * the ring buffer filling up (only if RB_FL_OVERWRITE is off).
4298 * @buffer: The ring buffer
4299 * @cpu: The per CPU buffer to get the number of overruns from
4302 ring_buffer_dropped_events_cpu(struct trace_buffer *buffer, int cpu)
4304 struct ring_buffer_per_cpu *cpu_buffer;
4307 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4310 cpu_buffer = buffer->buffers[cpu];
4311 ret = local_read(&cpu_buffer->dropped_events);
4315 EXPORT_SYMBOL_GPL(ring_buffer_dropped_events_cpu);
4318 * ring_buffer_read_events_cpu - get the number of events successfully read
4319 * @buffer: The ring buffer
4320 * @cpu: The per CPU buffer to get the number of events read
4323 ring_buffer_read_events_cpu(struct trace_buffer *buffer, int cpu)
4325 struct ring_buffer_per_cpu *cpu_buffer;
4327 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4330 cpu_buffer = buffer->buffers[cpu];
4331 return cpu_buffer->read;
4333 EXPORT_SYMBOL_GPL(ring_buffer_read_events_cpu);
4336 * ring_buffer_entries - get the number of entries in a buffer
4337 * @buffer: The ring buffer
4339 * Returns the total number of entries in the ring buffer
4342 unsigned long ring_buffer_entries(struct trace_buffer *buffer)
4344 struct ring_buffer_per_cpu *cpu_buffer;
4345 unsigned long entries = 0;
4348 /* if you care about this being correct, lock the buffer */
4349 for_each_buffer_cpu(buffer, cpu) {
4350 cpu_buffer = buffer->buffers[cpu];
4351 entries += rb_num_of_entries(cpu_buffer);
4356 EXPORT_SYMBOL_GPL(ring_buffer_entries);
4359 * ring_buffer_overruns - get the number of overruns in buffer
4360 * @buffer: The ring buffer
4362 * Returns the total number of overruns in the ring buffer
4365 unsigned long ring_buffer_overruns(struct trace_buffer *buffer)
4367 struct ring_buffer_per_cpu *cpu_buffer;
4368 unsigned long overruns = 0;
4371 /* if you care about this being correct, lock the buffer */
4372 for_each_buffer_cpu(buffer, cpu) {
4373 cpu_buffer = buffer->buffers[cpu];
4374 overruns += local_read(&cpu_buffer->overrun);
4379 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
4381 static void rb_iter_reset(struct ring_buffer_iter *iter)
4383 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
4385 /* Iterator usage is expected to have record disabled */
4386 iter->head_page = cpu_buffer->reader_page;
4387 iter->head = cpu_buffer->reader_page->read;
4388 iter->next_event = iter->head;
4390 iter->cache_reader_page = iter->head_page;
4391 iter->cache_read = cpu_buffer->read;
4392 iter->cache_pages_removed = cpu_buffer->pages_removed;
4395 iter->read_stamp = cpu_buffer->read_stamp;
4396 iter->page_stamp = cpu_buffer->reader_page->page->time_stamp;
4398 iter->read_stamp = iter->head_page->page->time_stamp;
4399 iter->page_stamp = iter->read_stamp;
4404 * ring_buffer_iter_reset - reset an iterator
4405 * @iter: The iterator to reset
4407 * Resets the iterator, so that it will start from the beginning
4410 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
4412 struct ring_buffer_per_cpu *cpu_buffer;
4413 unsigned long flags;
4418 cpu_buffer = iter->cpu_buffer;
4420 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4421 rb_iter_reset(iter);
4422 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4424 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
4427 * ring_buffer_iter_empty - check if an iterator has no more to read
4428 * @iter: The iterator to check
4430 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
4432 struct ring_buffer_per_cpu *cpu_buffer;
4433 struct buffer_page *reader;
4434 struct buffer_page *head_page;
4435 struct buffer_page *commit_page;
4436 struct buffer_page *curr_commit_page;
4441 cpu_buffer = iter->cpu_buffer;
4442 reader = cpu_buffer->reader_page;
4443 head_page = cpu_buffer->head_page;
4444 commit_page = cpu_buffer->commit_page;
4445 commit_ts = commit_page->page->time_stamp;
4448 * When the writer goes across pages, it issues a cmpxchg which
4449 * is a mb(), which will synchronize with the rmb here.
4450 * (see rb_tail_page_update())
4453 commit = rb_page_commit(commit_page);
4454 /* We want to make sure that the commit page doesn't change */
4457 /* Make sure commit page didn't change */
4458 curr_commit_page = READ_ONCE(cpu_buffer->commit_page);
4459 curr_commit_ts = READ_ONCE(curr_commit_page->page->time_stamp);
4461 /* If the commit page changed, then there's more data */
4462 if (curr_commit_page != commit_page ||
4463 curr_commit_ts != commit_ts)
4466 /* Still racy, as it may return a false positive, but that's OK */
4467 return ((iter->head_page == commit_page && iter->head >= commit) ||
4468 (iter->head_page == reader && commit_page == head_page &&
4469 head_page->read == commit &&
4470 iter->head == rb_page_commit(cpu_buffer->reader_page)));
4472 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
4475 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
4476 struct ring_buffer_event *event)
4480 switch (event->type_len) {
4481 case RINGBUF_TYPE_PADDING:
4484 case RINGBUF_TYPE_TIME_EXTEND:
4485 delta = rb_event_time_stamp(event);
4486 cpu_buffer->read_stamp += delta;
4489 case RINGBUF_TYPE_TIME_STAMP:
4490 delta = rb_event_time_stamp(event);
4491 delta = rb_fix_abs_ts(delta, cpu_buffer->read_stamp);
4492 cpu_buffer->read_stamp = delta;
4495 case RINGBUF_TYPE_DATA:
4496 cpu_buffer->read_stamp += event->time_delta;
4500 RB_WARN_ON(cpu_buffer, 1);
4505 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
4506 struct ring_buffer_event *event)
4510 switch (event->type_len) {
4511 case RINGBUF_TYPE_PADDING:
4514 case RINGBUF_TYPE_TIME_EXTEND:
4515 delta = rb_event_time_stamp(event);
4516 iter->read_stamp += delta;
4519 case RINGBUF_TYPE_TIME_STAMP:
4520 delta = rb_event_time_stamp(event);
4521 delta = rb_fix_abs_ts(delta, iter->read_stamp);
4522 iter->read_stamp = delta;
4525 case RINGBUF_TYPE_DATA:
4526 iter->read_stamp += event->time_delta;
4530 RB_WARN_ON(iter->cpu_buffer, 1);
4534 static struct buffer_page *
4535 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
4537 struct buffer_page *reader = NULL;
4538 unsigned long overwrite;
4539 unsigned long flags;
4543 local_irq_save(flags);
4544 arch_spin_lock(&cpu_buffer->lock);
4548 * This should normally only loop twice. But because the
4549 * start of the reader inserts an empty page, it causes
4550 * a case where we will loop three times. There should be no
4551 * reason to loop four times (that I know of).
4553 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
4558 reader = cpu_buffer->reader_page;
4560 /* If there's more to read, return this page */
4561 if (cpu_buffer->reader_page->read < rb_page_size(reader))
4564 /* Never should we have an index greater than the size */
4565 if (RB_WARN_ON(cpu_buffer,
4566 cpu_buffer->reader_page->read > rb_page_size(reader)))
4569 /* check if we caught up to the tail */
4571 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
4574 /* Don't bother swapping if the ring buffer is empty */
4575 if (rb_num_of_entries(cpu_buffer) == 0)
4579 * Reset the reader page to size zero.
4581 local_set(&cpu_buffer->reader_page->write, 0);
4582 local_set(&cpu_buffer->reader_page->entries, 0);
4583 local_set(&cpu_buffer->reader_page->page->commit, 0);
4584 cpu_buffer->reader_page->real_end = 0;
4588 * Splice the empty reader page into the list around the head.
4590 reader = rb_set_head_page(cpu_buffer);
4593 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
4594 cpu_buffer->reader_page->list.prev = reader->list.prev;
4597 * cpu_buffer->pages just needs to point to the buffer, it
4598 * has no specific buffer page to point to. Lets move it out
4599 * of our way so we don't accidentally swap it.
4601 cpu_buffer->pages = reader->list.prev;
4603 /* The reader page will be pointing to the new head */
4604 rb_set_list_to_head(&cpu_buffer->reader_page->list);
4607 * We want to make sure we read the overruns after we set up our
4608 * pointers to the next object. The writer side does a
4609 * cmpxchg to cross pages which acts as the mb on the writer
4610 * side. Note, the reader will constantly fail the swap
4611 * while the writer is updating the pointers, so this
4612 * guarantees that the overwrite recorded here is the one we
4613 * want to compare with the last_overrun.
4616 overwrite = local_read(&(cpu_buffer->overrun));
4619 * Here's the tricky part.
4621 * We need to move the pointer past the header page.
4622 * But we can only do that if a writer is not currently
4623 * moving it. The page before the header page has the
4624 * flag bit '1' set if it is pointing to the page we want.
4625 * but if the writer is in the process of moving it
4626 * than it will be '2' or already moved '0'.
4629 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
4632 * If we did not convert it, then we must try again.
4638 * Yay! We succeeded in replacing the page.
4640 * Now make the new head point back to the reader page.
4642 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
4643 rb_inc_page(&cpu_buffer->head_page);
4645 local_inc(&cpu_buffer->pages_read);
4647 /* Finally update the reader page to the new head */
4648 cpu_buffer->reader_page = reader;
4649 cpu_buffer->reader_page->read = 0;
4651 if (overwrite != cpu_buffer->last_overrun) {
4652 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
4653 cpu_buffer->last_overrun = overwrite;
4659 /* Update the read_stamp on the first event */
4660 if (reader && reader->read == 0)
4661 cpu_buffer->read_stamp = reader->page->time_stamp;
4663 arch_spin_unlock(&cpu_buffer->lock);
4664 local_irq_restore(flags);
4667 * The writer has preempt disable, wait for it. But not forever
4668 * Although, 1 second is pretty much "forever"
4670 #define USECS_WAIT 1000000
4671 for (nr_loops = 0; nr_loops < USECS_WAIT; nr_loops++) {
4672 /* If the write is past the end of page, a writer is still updating it */
4673 if (likely(!reader || rb_page_write(reader) <= BUF_PAGE_SIZE))
4678 /* Get the latest version of the reader write value */
4682 /* The writer is not moving forward? Something is wrong */
4683 if (RB_WARN_ON(cpu_buffer, nr_loops == USECS_WAIT))
4687 * Make sure we see any padding after the write update
4688 * (see rb_reset_tail()).
4690 * In addition, a writer may be writing on the reader page
4691 * if the page has not been fully filled, so the read barrier
4692 * is also needed to make sure we see the content of what is
4693 * committed by the writer (see rb_set_commit_to_write()).
4701 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
4703 struct ring_buffer_event *event;
4704 struct buffer_page *reader;
4707 reader = rb_get_reader_page(cpu_buffer);
4709 /* This function should not be called when buffer is empty */
4710 if (RB_WARN_ON(cpu_buffer, !reader))
4713 event = rb_reader_event(cpu_buffer);
4715 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
4718 rb_update_read_stamp(cpu_buffer, event);
4720 length = rb_event_length(event);
4721 cpu_buffer->reader_page->read += length;
4722 cpu_buffer->read_bytes += length;
4725 static void rb_advance_iter(struct ring_buffer_iter *iter)
4727 struct ring_buffer_per_cpu *cpu_buffer;
4729 cpu_buffer = iter->cpu_buffer;
4731 /* If head == next_event then we need to jump to the next event */
4732 if (iter->head == iter->next_event) {
4733 /* If the event gets overwritten again, there's nothing to do */
4734 if (rb_iter_head_event(iter) == NULL)
4738 iter->head = iter->next_event;
4741 * Check if we are at the end of the buffer.
4743 if (iter->next_event >= rb_page_size(iter->head_page)) {
4744 /* discarded commits can make the page empty */
4745 if (iter->head_page == cpu_buffer->commit_page)
4751 rb_update_iter_read_stamp(iter, iter->event);
4754 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
4756 return cpu_buffer->lost_events;
4759 static struct ring_buffer_event *
4760 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
4761 unsigned long *lost_events)
4763 struct ring_buffer_event *event;
4764 struct buffer_page *reader;
4771 * We repeat when a time extend is encountered.
4772 * Since the time extend is always attached to a data event,
4773 * we should never loop more than once.
4774 * (We never hit the following condition more than twice).
4776 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
4779 reader = rb_get_reader_page(cpu_buffer);
4783 event = rb_reader_event(cpu_buffer);
4785 switch (event->type_len) {
4786 case RINGBUF_TYPE_PADDING:
4787 if (rb_null_event(event))
4788 RB_WARN_ON(cpu_buffer, 1);
4790 * Because the writer could be discarding every
4791 * event it creates (which would probably be bad)
4792 * if we were to go back to "again" then we may never
4793 * catch up, and will trigger the warn on, or lock
4794 * the box. Return the padding, and we will release
4795 * the current locks, and try again.
4799 case RINGBUF_TYPE_TIME_EXTEND:
4800 /* Internal data, OK to advance */
4801 rb_advance_reader(cpu_buffer);
4804 case RINGBUF_TYPE_TIME_STAMP:
4806 *ts = rb_event_time_stamp(event);
4807 *ts = rb_fix_abs_ts(*ts, reader->page->time_stamp);
4808 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4809 cpu_buffer->cpu, ts);
4811 /* Internal data, OK to advance */
4812 rb_advance_reader(cpu_buffer);
4815 case RINGBUF_TYPE_DATA:
4817 *ts = cpu_buffer->read_stamp + event->time_delta;
4818 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4819 cpu_buffer->cpu, ts);
4822 *lost_events = rb_lost_events(cpu_buffer);
4826 RB_WARN_ON(cpu_buffer, 1);
4831 EXPORT_SYMBOL_GPL(ring_buffer_peek);
4833 static struct ring_buffer_event *
4834 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
4836 struct trace_buffer *buffer;
4837 struct ring_buffer_per_cpu *cpu_buffer;
4838 struct ring_buffer_event *event;
4844 cpu_buffer = iter->cpu_buffer;
4845 buffer = cpu_buffer->buffer;
4848 * Check if someone performed a consuming read to the buffer
4849 * or removed some pages from the buffer. In these cases,
4850 * iterator was invalidated and we need to reset it.
4852 if (unlikely(iter->cache_read != cpu_buffer->read ||
4853 iter->cache_reader_page != cpu_buffer->reader_page ||
4854 iter->cache_pages_removed != cpu_buffer->pages_removed))
4855 rb_iter_reset(iter);
4858 if (ring_buffer_iter_empty(iter))
4862 * As the writer can mess with what the iterator is trying
4863 * to read, just give up if we fail to get an event after
4864 * three tries. The iterator is not as reliable when reading
4865 * the ring buffer with an active write as the consumer is.
4866 * Do not warn if the three failures is reached.
4871 if (rb_per_cpu_empty(cpu_buffer))
4874 if (iter->head >= rb_page_size(iter->head_page)) {
4879 event = rb_iter_head_event(iter);
4883 switch (event->type_len) {
4884 case RINGBUF_TYPE_PADDING:
4885 if (rb_null_event(event)) {
4889 rb_advance_iter(iter);
4892 case RINGBUF_TYPE_TIME_EXTEND:
4893 /* Internal data, OK to advance */
4894 rb_advance_iter(iter);
4897 case RINGBUF_TYPE_TIME_STAMP:
4899 *ts = rb_event_time_stamp(event);
4900 *ts = rb_fix_abs_ts(*ts, iter->head_page->page->time_stamp);
4901 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
4902 cpu_buffer->cpu, ts);
4904 /* Internal data, OK to advance */
4905 rb_advance_iter(iter);
4908 case RINGBUF_TYPE_DATA:
4910 *ts = iter->read_stamp + event->time_delta;
4911 ring_buffer_normalize_time_stamp(buffer,
4912 cpu_buffer->cpu, ts);
4917 RB_WARN_ON(cpu_buffer, 1);
4922 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
4924 static inline bool rb_reader_lock(struct ring_buffer_per_cpu *cpu_buffer)
4926 if (likely(!in_nmi())) {
4927 raw_spin_lock(&cpu_buffer->reader_lock);
4932 * If an NMI die dumps out the content of the ring buffer
4933 * trylock must be used to prevent a deadlock if the NMI
4934 * preempted a task that holds the ring buffer locks. If
4935 * we get the lock then all is fine, if not, then continue
4936 * to do the read, but this can corrupt the ring buffer,
4937 * so it must be permanently disabled from future writes.
4938 * Reading from NMI is a oneshot deal.
4940 if (raw_spin_trylock(&cpu_buffer->reader_lock))
4943 /* Continue without locking, but disable the ring buffer */
4944 atomic_inc(&cpu_buffer->record_disabled);
4949 rb_reader_unlock(struct ring_buffer_per_cpu *cpu_buffer, bool locked)
4952 raw_spin_unlock(&cpu_buffer->reader_lock);
4956 * ring_buffer_peek - peek at the next event to be read
4957 * @buffer: The ring buffer to read
4958 * @cpu: The cpu to peak at
4959 * @ts: The timestamp counter of this event.
4960 * @lost_events: a variable to store if events were lost (may be NULL)
4962 * This will return the event that will be read next, but does
4963 * not consume the data.
4965 struct ring_buffer_event *
4966 ring_buffer_peek(struct trace_buffer *buffer, int cpu, u64 *ts,
4967 unsigned long *lost_events)
4969 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4970 struct ring_buffer_event *event;
4971 unsigned long flags;
4974 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4978 local_irq_save(flags);
4979 dolock = rb_reader_lock(cpu_buffer);
4980 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
4981 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4982 rb_advance_reader(cpu_buffer);
4983 rb_reader_unlock(cpu_buffer, dolock);
4984 local_irq_restore(flags);
4986 if (event && event->type_len == RINGBUF_TYPE_PADDING)
4992 /** ring_buffer_iter_dropped - report if there are dropped events
4993 * @iter: The ring buffer iterator
4995 * Returns true if there was dropped events since the last peek.
4997 bool ring_buffer_iter_dropped(struct ring_buffer_iter *iter)
4999 bool ret = iter->missed_events != 0;
5001 iter->missed_events = 0;
5004 EXPORT_SYMBOL_GPL(ring_buffer_iter_dropped);
5007 * ring_buffer_iter_peek - peek at the next event to be read
5008 * @iter: The ring buffer iterator
5009 * @ts: The timestamp counter of this event.
5011 * This will return the event that will be read next, but does
5012 * not increment the iterator.
5014 struct ring_buffer_event *
5015 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
5017 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5018 struct ring_buffer_event *event;
5019 unsigned long flags;
5022 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5023 event = rb_iter_peek(iter, ts);
5024 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5026 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5033 * ring_buffer_consume - return an event and consume it
5034 * @buffer: The ring buffer to get the next event from
5035 * @cpu: the cpu to read the buffer from
5036 * @ts: a variable to store the timestamp (may be NULL)
5037 * @lost_events: a variable to store if events were lost (may be NULL)
5039 * Returns the next event in the ring buffer, and that event is consumed.
5040 * Meaning, that sequential reads will keep returning a different event,
5041 * and eventually empty the ring buffer if the producer is slower.
5043 struct ring_buffer_event *
5044 ring_buffer_consume(struct trace_buffer *buffer, int cpu, u64 *ts,
5045 unsigned long *lost_events)
5047 struct ring_buffer_per_cpu *cpu_buffer;
5048 struct ring_buffer_event *event = NULL;
5049 unsigned long flags;
5053 /* might be called in atomic */
5056 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5059 cpu_buffer = buffer->buffers[cpu];
5060 local_irq_save(flags);
5061 dolock = rb_reader_lock(cpu_buffer);
5063 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
5065 cpu_buffer->lost_events = 0;
5066 rb_advance_reader(cpu_buffer);
5069 rb_reader_unlock(cpu_buffer, dolock);
5070 local_irq_restore(flags);
5075 if (event && event->type_len == RINGBUF_TYPE_PADDING)
5080 EXPORT_SYMBOL_GPL(ring_buffer_consume);
5083 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
5084 * @buffer: The ring buffer to read from
5085 * @cpu: The cpu buffer to iterate over
5086 * @flags: gfp flags to use for memory allocation
5088 * This performs the initial preparations necessary to iterate
5089 * through the buffer. Memory is allocated, buffer recording
5090 * is disabled, and the iterator pointer is returned to the caller.
5092 * Disabling buffer recording prevents the reading from being
5093 * corrupted. This is not a consuming read, so a producer is not
5096 * After a sequence of ring_buffer_read_prepare calls, the user is
5097 * expected to make at least one call to ring_buffer_read_prepare_sync.
5098 * Afterwards, ring_buffer_read_start is invoked to get things going
5101 * This overall must be paired with ring_buffer_read_finish.
5103 struct ring_buffer_iter *
5104 ring_buffer_read_prepare(struct trace_buffer *buffer, int cpu, gfp_t flags)
5106 struct ring_buffer_per_cpu *cpu_buffer;
5107 struct ring_buffer_iter *iter;
5109 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5112 iter = kzalloc(sizeof(*iter), flags);
5116 iter->event = kmalloc(BUF_MAX_DATA_SIZE, flags);
5122 cpu_buffer = buffer->buffers[cpu];
5124 iter->cpu_buffer = cpu_buffer;
5126 atomic_inc(&cpu_buffer->resize_disabled);
5130 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
5133 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
5135 * All previously invoked ring_buffer_read_prepare calls to prepare
5136 * iterators will be synchronized. Afterwards, read_buffer_read_start
5137 * calls on those iterators are allowed.
5140 ring_buffer_read_prepare_sync(void)
5144 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
5147 * ring_buffer_read_start - start a non consuming read of the buffer
5148 * @iter: The iterator returned by ring_buffer_read_prepare
5150 * This finalizes the startup of an iteration through the buffer.
5151 * The iterator comes from a call to ring_buffer_read_prepare and
5152 * an intervening ring_buffer_read_prepare_sync must have been
5155 * Must be paired with ring_buffer_read_finish.
5158 ring_buffer_read_start(struct ring_buffer_iter *iter)
5160 struct ring_buffer_per_cpu *cpu_buffer;
5161 unsigned long flags;
5166 cpu_buffer = iter->cpu_buffer;
5168 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5169 arch_spin_lock(&cpu_buffer->lock);
5170 rb_iter_reset(iter);
5171 arch_spin_unlock(&cpu_buffer->lock);
5172 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5174 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
5177 * ring_buffer_read_finish - finish reading the iterator of the buffer
5178 * @iter: The iterator retrieved by ring_buffer_start
5180 * This re-enables the recording to the buffer, and frees the
5184 ring_buffer_read_finish(struct ring_buffer_iter *iter)
5186 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5187 unsigned long flags;
5190 * Ring buffer is disabled from recording, here's a good place
5191 * to check the integrity of the ring buffer.
5192 * Must prevent readers from trying to read, as the check
5193 * clears the HEAD page and readers require it.
5195 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5196 rb_check_pages(cpu_buffer);
5197 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5199 atomic_dec(&cpu_buffer->resize_disabled);
5203 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
5206 * ring_buffer_iter_advance - advance the iterator to the next location
5207 * @iter: The ring buffer iterator
5209 * Move the location of the iterator such that the next read will
5210 * be the next location of the iterator.
5212 void ring_buffer_iter_advance(struct ring_buffer_iter *iter)
5214 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
5215 unsigned long flags;
5217 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5219 rb_advance_iter(iter);
5221 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5223 EXPORT_SYMBOL_GPL(ring_buffer_iter_advance);
5226 * ring_buffer_size - return the size of the ring buffer (in bytes)
5227 * @buffer: The ring buffer.
5228 * @cpu: The CPU to get ring buffer size from.
5230 unsigned long ring_buffer_size(struct trace_buffer *buffer, int cpu)
5233 * Earlier, this method returned
5234 * BUF_PAGE_SIZE * buffer->nr_pages
5235 * Since the nr_pages field is now removed, we have converted this to
5236 * return the per cpu buffer value.
5238 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5241 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
5243 EXPORT_SYMBOL_GPL(ring_buffer_size);
5245 static void rb_clear_buffer_page(struct buffer_page *page)
5247 local_set(&page->write, 0);
5248 local_set(&page->entries, 0);
5249 rb_init_page(page->page);
5254 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
5256 struct buffer_page *page;
5258 rb_head_page_deactivate(cpu_buffer);
5260 cpu_buffer->head_page
5261 = list_entry(cpu_buffer->pages, struct buffer_page, list);
5262 rb_clear_buffer_page(cpu_buffer->head_page);
5263 list_for_each_entry(page, cpu_buffer->pages, list) {
5264 rb_clear_buffer_page(page);
5267 cpu_buffer->tail_page = cpu_buffer->head_page;
5268 cpu_buffer->commit_page = cpu_buffer->head_page;
5270 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
5271 INIT_LIST_HEAD(&cpu_buffer->new_pages);
5272 rb_clear_buffer_page(cpu_buffer->reader_page);
5274 local_set(&cpu_buffer->entries_bytes, 0);
5275 local_set(&cpu_buffer->overrun, 0);
5276 local_set(&cpu_buffer->commit_overrun, 0);
5277 local_set(&cpu_buffer->dropped_events, 0);
5278 local_set(&cpu_buffer->entries, 0);
5279 local_set(&cpu_buffer->committing, 0);
5280 local_set(&cpu_buffer->commits, 0);
5281 local_set(&cpu_buffer->pages_touched, 0);
5282 local_set(&cpu_buffer->pages_lost, 0);
5283 local_set(&cpu_buffer->pages_read, 0);
5284 cpu_buffer->last_pages_touch = 0;
5285 cpu_buffer->shortest_full = 0;
5286 cpu_buffer->read = 0;
5287 cpu_buffer->read_bytes = 0;
5289 rb_time_set(&cpu_buffer->write_stamp, 0);
5290 rb_time_set(&cpu_buffer->before_stamp, 0);
5292 memset(cpu_buffer->event_stamp, 0, sizeof(cpu_buffer->event_stamp));
5294 cpu_buffer->lost_events = 0;
5295 cpu_buffer->last_overrun = 0;
5297 rb_head_page_activate(cpu_buffer);
5298 cpu_buffer->pages_removed = 0;
5301 /* Must have disabled the cpu buffer then done a synchronize_rcu */
5302 static void reset_disabled_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
5304 unsigned long flags;
5306 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5308 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
5311 arch_spin_lock(&cpu_buffer->lock);
5313 rb_reset_cpu(cpu_buffer);
5315 arch_spin_unlock(&cpu_buffer->lock);
5318 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5322 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
5323 * @buffer: The ring buffer to reset a per cpu buffer of
5324 * @cpu: The CPU buffer to be reset
5326 void ring_buffer_reset_cpu(struct trace_buffer *buffer, int cpu)
5328 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5330 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5333 /* prevent another thread from changing buffer sizes */
5334 mutex_lock(&buffer->mutex);
5336 atomic_inc(&cpu_buffer->resize_disabled);
5337 atomic_inc(&cpu_buffer->record_disabled);
5339 /* Make sure all commits have finished */
5342 reset_disabled_cpu_buffer(cpu_buffer);
5344 atomic_dec(&cpu_buffer->record_disabled);
5345 atomic_dec(&cpu_buffer->resize_disabled);
5347 mutex_unlock(&buffer->mutex);
5349 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
5351 /* Flag to ensure proper resetting of atomic variables */
5352 #define RESET_BIT (1 << 30)
5355 * ring_buffer_reset_online_cpus - reset a ring buffer per CPU buffer
5356 * @buffer: The ring buffer to reset a per cpu buffer of
5358 void ring_buffer_reset_online_cpus(struct trace_buffer *buffer)
5360 struct ring_buffer_per_cpu *cpu_buffer;
5363 /* prevent another thread from changing buffer sizes */
5364 mutex_lock(&buffer->mutex);
5366 for_each_online_buffer_cpu(buffer, cpu) {
5367 cpu_buffer = buffer->buffers[cpu];
5369 atomic_add(RESET_BIT, &cpu_buffer->resize_disabled);
5370 atomic_inc(&cpu_buffer->record_disabled);
5373 /* Make sure all commits have finished */
5376 for_each_buffer_cpu(buffer, cpu) {
5377 cpu_buffer = buffer->buffers[cpu];
5380 * If a CPU came online during the synchronize_rcu(), then
5383 if (!(atomic_read(&cpu_buffer->resize_disabled) & RESET_BIT))
5386 reset_disabled_cpu_buffer(cpu_buffer);
5388 atomic_dec(&cpu_buffer->record_disabled);
5389 atomic_sub(RESET_BIT, &cpu_buffer->resize_disabled);
5392 mutex_unlock(&buffer->mutex);
5396 * ring_buffer_reset - reset a ring buffer
5397 * @buffer: The ring buffer to reset all cpu buffers
5399 void ring_buffer_reset(struct trace_buffer *buffer)
5401 struct ring_buffer_per_cpu *cpu_buffer;
5404 /* prevent another thread from changing buffer sizes */
5405 mutex_lock(&buffer->mutex);
5407 for_each_buffer_cpu(buffer, cpu) {
5408 cpu_buffer = buffer->buffers[cpu];
5410 atomic_inc(&cpu_buffer->resize_disabled);
5411 atomic_inc(&cpu_buffer->record_disabled);
5414 /* Make sure all commits have finished */
5417 for_each_buffer_cpu(buffer, cpu) {
5418 cpu_buffer = buffer->buffers[cpu];
5420 reset_disabled_cpu_buffer(cpu_buffer);
5422 atomic_dec(&cpu_buffer->record_disabled);
5423 atomic_dec(&cpu_buffer->resize_disabled);
5426 mutex_unlock(&buffer->mutex);
5428 EXPORT_SYMBOL_GPL(ring_buffer_reset);
5431 * ring_buffer_empty - is the ring buffer empty?
5432 * @buffer: The ring buffer to test
5434 bool ring_buffer_empty(struct trace_buffer *buffer)
5436 struct ring_buffer_per_cpu *cpu_buffer;
5437 unsigned long flags;
5442 /* yes this is racy, but if you don't like the race, lock the buffer */
5443 for_each_buffer_cpu(buffer, cpu) {
5444 cpu_buffer = buffer->buffers[cpu];
5445 local_irq_save(flags);
5446 dolock = rb_reader_lock(cpu_buffer);
5447 ret = rb_per_cpu_empty(cpu_buffer);
5448 rb_reader_unlock(cpu_buffer, dolock);
5449 local_irq_restore(flags);
5457 EXPORT_SYMBOL_GPL(ring_buffer_empty);
5460 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
5461 * @buffer: The ring buffer
5462 * @cpu: The CPU buffer to test
5464 bool ring_buffer_empty_cpu(struct trace_buffer *buffer, int cpu)
5466 struct ring_buffer_per_cpu *cpu_buffer;
5467 unsigned long flags;
5471 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5474 cpu_buffer = buffer->buffers[cpu];
5475 local_irq_save(flags);
5476 dolock = rb_reader_lock(cpu_buffer);
5477 ret = rb_per_cpu_empty(cpu_buffer);
5478 rb_reader_unlock(cpu_buffer, dolock);
5479 local_irq_restore(flags);
5483 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
5485 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
5487 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
5488 * @buffer_a: One buffer to swap with
5489 * @buffer_b: The other buffer to swap with
5490 * @cpu: the CPU of the buffers to swap
5492 * This function is useful for tracers that want to take a "snapshot"
5493 * of a CPU buffer and has another back up buffer lying around.
5494 * it is expected that the tracer handles the cpu buffer not being
5495 * used at the moment.
5497 int ring_buffer_swap_cpu(struct trace_buffer *buffer_a,
5498 struct trace_buffer *buffer_b, int cpu)
5500 struct ring_buffer_per_cpu *cpu_buffer_a;
5501 struct ring_buffer_per_cpu *cpu_buffer_b;
5504 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
5505 !cpumask_test_cpu(cpu, buffer_b->cpumask))
5508 cpu_buffer_a = buffer_a->buffers[cpu];
5509 cpu_buffer_b = buffer_b->buffers[cpu];
5511 /* At least make sure the two buffers are somewhat the same */
5512 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
5517 if (atomic_read(&buffer_a->record_disabled))
5520 if (atomic_read(&buffer_b->record_disabled))
5523 if (atomic_read(&cpu_buffer_a->record_disabled))
5526 if (atomic_read(&cpu_buffer_b->record_disabled))
5530 * We can't do a synchronize_rcu here because this
5531 * function can be called in atomic context.
5532 * Normally this will be called from the same CPU as cpu.
5533 * If not it's up to the caller to protect this.
5535 atomic_inc(&cpu_buffer_a->record_disabled);
5536 atomic_inc(&cpu_buffer_b->record_disabled);
5539 if (local_read(&cpu_buffer_a->committing))
5541 if (local_read(&cpu_buffer_b->committing))
5545 * When resize is in progress, we cannot swap it because
5546 * it will mess the state of the cpu buffer.
5548 if (atomic_read(&buffer_a->resizing))
5550 if (atomic_read(&buffer_b->resizing))
5553 buffer_a->buffers[cpu] = cpu_buffer_b;
5554 buffer_b->buffers[cpu] = cpu_buffer_a;
5556 cpu_buffer_b->buffer = buffer_a;
5557 cpu_buffer_a->buffer = buffer_b;
5562 atomic_dec(&cpu_buffer_a->record_disabled);
5563 atomic_dec(&cpu_buffer_b->record_disabled);
5567 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
5568 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
5571 * ring_buffer_alloc_read_page - allocate a page to read from buffer
5572 * @buffer: the buffer to allocate for.
5573 * @cpu: the cpu buffer to allocate.
5575 * This function is used in conjunction with ring_buffer_read_page.
5576 * When reading a full page from the ring buffer, these functions
5577 * can be used to speed up the process. The calling function should
5578 * allocate a few pages first with this function. Then when it
5579 * needs to get pages from the ring buffer, it passes the result
5580 * of this function into ring_buffer_read_page, which will swap
5581 * the page that was allocated, with the read page of the buffer.
5584 * The page allocated, or ERR_PTR
5586 void *ring_buffer_alloc_read_page(struct trace_buffer *buffer, int cpu)
5588 struct ring_buffer_per_cpu *cpu_buffer;
5589 struct buffer_data_page *bpage = NULL;
5590 unsigned long flags;
5593 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5594 return ERR_PTR(-ENODEV);
5596 cpu_buffer = buffer->buffers[cpu];
5597 local_irq_save(flags);
5598 arch_spin_lock(&cpu_buffer->lock);
5600 if (cpu_buffer->free_page) {
5601 bpage = cpu_buffer->free_page;
5602 cpu_buffer->free_page = NULL;
5605 arch_spin_unlock(&cpu_buffer->lock);
5606 local_irq_restore(flags);
5611 page = alloc_pages_node(cpu_to_node(cpu),
5612 GFP_KERNEL | __GFP_NORETRY, 0);
5614 return ERR_PTR(-ENOMEM);
5616 bpage = page_address(page);
5619 rb_init_page(bpage);
5623 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
5626 * ring_buffer_free_read_page - free an allocated read page
5627 * @buffer: the buffer the page was allocate for
5628 * @cpu: the cpu buffer the page came from
5629 * @data: the page to free
5631 * Free a page allocated from ring_buffer_alloc_read_page.
5633 void ring_buffer_free_read_page(struct trace_buffer *buffer, int cpu, void *data)
5635 struct ring_buffer_per_cpu *cpu_buffer;
5636 struct buffer_data_page *bpage = data;
5637 struct page *page = virt_to_page(bpage);
5638 unsigned long flags;
5640 if (!buffer || !buffer->buffers || !buffer->buffers[cpu])
5643 cpu_buffer = buffer->buffers[cpu];
5645 /* If the page is still in use someplace else, we can't reuse it */
5646 if (page_ref_count(page) > 1)
5649 local_irq_save(flags);
5650 arch_spin_lock(&cpu_buffer->lock);
5652 if (!cpu_buffer->free_page) {
5653 cpu_buffer->free_page = bpage;
5657 arch_spin_unlock(&cpu_buffer->lock);
5658 local_irq_restore(flags);
5661 free_page((unsigned long)bpage);
5663 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
5666 * ring_buffer_read_page - extract a page from the ring buffer
5667 * @buffer: buffer to extract from
5668 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
5669 * @len: amount to extract
5670 * @cpu: the cpu of the buffer to extract
5671 * @full: should the extraction only happen when the page is full.
5673 * This function will pull out a page from the ring buffer and consume it.
5674 * @data_page must be the address of the variable that was returned
5675 * from ring_buffer_alloc_read_page. This is because the page might be used
5676 * to swap with a page in the ring buffer.
5679 * rpage = ring_buffer_alloc_read_page(buffer, cpu);
5680 * if (IS_ERR(rpage))
5681 * return PTR_ERR(rpage);
5682 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
5684 * process_page(rpage, ret);
5686 * When @full is set, the function will not return true unless
5687 * the writer is off the reader page.
5689 * Note: it is up to the calling functions to handle sleeps and wakeups.
5690 * The ring buffer can be used anywhere in the kernel and can not
5691 * blindly call wake_up. The layer that uses the ring buffer must be
5692 * responsible for that.
5695 * >=0 if data has been transferred, returns the offset of consumed data.
5696 * <0 if no data has been transferred.
5698 int ring_buffer_read_page(struct trace_buffer *buffer,
5699 void **data_page, size_t len, int cpu, int full)
5701 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
5702 struct ring_buffer_event *event;
5703 struct buffer_data_page *bpage;
5704 struct buffer_page *reader;
5705 unsigned long missed_events;
5706 unsigned long flags;
5707 unsigned int commit;
5712 if (!cpumask_test_cpu(cpu, buffer->cpumask))
5716 * If len is not big enough to hold the page header, then
5717 * we can not copy anything.
5719 if (len <= BUF_PAGE_HDR_SIZE)
5722 len -= BUF_PAGE_HDR_SIZE;
5731 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
5733 reader = rb_get_reader_page(cpu_buffer);
5737 event = rb_reader_event(cpu_buffer);
5739 read = reader->read;
5740 commit = rb_page_commit(reader);
5742 /* Check if any events were dropped */
5743 missed_events = cpu_buffer->lost_events;
5746 * If this page has been partially read or
5747 * if len is not big enough to read the rest of the page or
5748 * a writer is still on the page, then
5749 * we must copy the data from the page to the buffer.
5750 * Otherwise, we can simply swap the page with the one passed in.
5752 if (read || (len < (commit - read)) ||
5753 cpu_buffer->reader_page == cpu_buffer->commit_page) {
5754 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
5755 unsigned int rpos = read;
5756 unsigned int pos = 0;
5760 * If a full page is expected, this can still be returned
5761 * if there's been a previous partial read and the
5762 * rest of the page can be read and the commit page is off
5766 (!read || (len < (commit - read)) ||
5767 cpu_buffer->reader_page == cpu_buffer->commit_page))
5770 if (len > (commit - read))
5771 len = (commit - read);
5773 /* Always keep the time extend and data together */
5774 size = rb_event_ts_length(event);
5779 /* save the current timestamp, since the user will need it */
5780 save_timestamp = cpu_buffer->read_stamp;
5782 /* Need to copy one event at a time */
5784 /* We need the size of one event, because
5785 * rb_advance_reader only advances by one event,
5786 * whereas rb_event_ts_length may include the size of
5787 * one or two events.
5788 * We have already ensured there's enough space if this
5789 * is a time extend. */
5790 size = rb_event_length(event);
5791 memcpy(bpage->data + pos, rpage->data + rpos, size);
5795 rb_advance_reader(cpu_buffer);
5796 rpos = reader->read;
5802 event = rb_reader_event(cpu_buffer);
5803 /* Always keep the time extend and data together */
5804 size = rb_event_ts_length(event);
5805 } while (len >= size);
5808 local_set(&bpage->commit, pos);
5809 bpage->time_stamp = save_timestamp;
5811 /* we copied everything to the beginning */
5814 /* update the entry counter */
5815 cpu_buffer->read += rb_page_entries(reader);
5816 cpu_buffer->read_bytes += rb_page_commit(reader);
5818 /* swap the pages */
5819 rb_init_page(bpage);
5820 bpage = reader->page;
5821 reader->page = *data_page;
5822 local_set(&reader->write, 0);
5823 local_set(&reader->entries, 0);
5828 * Use the real_end for the data size,
5829 * This gives us a chance to store the lost events
5832 if (reader->real_end)
5833 local_set(&bpage->commit, reader->real_end);
5837 cpu_buffer->lost_events = 0;
5839 commit = local_read(&bpage->commit);
5841 * Set a flag in the commit field if we lost events
5843 if (missed_events) {
5844 /* If there is room at the end of the page to save the
5845 * missed events, then record it there.
5847 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
5848 memcpy(&bpage->data[commit], &missed_events,
5849 sizeof(missed_events));
5850 local_add(RB_MISSED_STORED, &bpage->commit);
5851 commit += sizeof(missed_events);
5853 local_add(RB_MISSED_EVENTS, &bpage->commit);
5857 * This page may be off to user land. Zero it out here.
5859 if (commit < BUF_PAGE_SIZE)
5860 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
5863 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
5868 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
5871 * We only allocate new buffers, never free them if the CPU goes down.
5872 * If we were to free the buffer, then the user would lose any trace that was in
5875 int trace_rb_cpu_prepare(unsigned int cpu, struct hlist_node *node)
5877 struct trace_buffer *buffer;
5880 unsigned long nr_pages;
5882 buffer = container_of(node, struct trace_buffer, node);
5883 if (cpumask_test_cpu(cpu, buffer->cpumask))
5888 /* check if all cpu sizes are same */
5889 for_each_buffer_cpu(buffer, cpu_i) {
5890 /* fill in the size from first enabled cpu */
5892 nr_pages = buffer->buffers[cpu_i]->nr_pages;
5893 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
5898 /* allocate minimum pages, user can later expand it */
5901 buffer->buffers[cpu] =
5902 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
5903 if (!buffer->buffers[cpu]) {
5904 WARN(1, "failed to allocate ring buffer on CPU %u\n",
5909 cpumask_set_cpu(cpu, buffer->cpumask);
5913 #ifdef CONFIG_RING_BUFFER_STARTUP_TEST
5915 * This is a basic integrity check of the ring buffer.
5916 * Late in the boot cycle this test will run when configured in.
5917 * It will kick off a thread per CPU that will go into a loop
5918 * writing to the per cpu ring buffer various sizes of data.
5919 * Some of the data will be large items, some small.
5921 * Another thread is created that goes into a spin, sending out
5922 * IPIs to the other CPUs to also write into the ring buffer.
5923 * this is to test the nesting ability of the buffer.
5925 * Basic stats are recorded and reported. If something in the
5926 * ring buffer should happen that's not expected, a big warning
5927 * is displayed and all ring buffers are disabled.
5929 static struct task_struct *rb_threads[NR_CPUS] __initdata;
5931 struct rb_test_data {
5932 struct trace_buffer *buffer;
5933 unsigned long events;
5934 unsigned long bytes_written;
5935 unsigned long bytes_alloc;
5936 unsigned long bytes_dropped;
5937 unsigned long events_nested;
5938 unsigned long bytes_written_nested;
5939 unsigned long bytes_alloc_nested;
5940 unsigned long bytes_dropped_nested;
5941 int min_size_nested;
5942 int max_size_nested;
5949 static struct rb_test_data rb_data[NR_CPUS] __initdata;
5952 #define RB_TEST_BUFFER_SIZE 1048576
5954 static char rb_string[] __initdata =
5955 "abcdefghijklmnopqrstuvwxyz1234567890!@#$%^&*()?+\\"
5956 "?+|:';\",.<>/?abcdefghijklmnopqrstuvwxyz1234567890"
5957 "!@#$%^&*()?+\\?+|:';\",.<>/?abcdefghijklmnopqrstuv";
5959 static bool rb_test_started __initdata;
5966 static __init int rb_write_something(struct rb_test_data *data, bool nested)
5968 struct ring_buffer_event *event;
5969 struct rb_item *item;
5976 /* Have nested writes different that what is written */
5977 cnt = data->cnt + (nested ? 27 : 0);
5979 /* Multiply cnt by ~e, to make some unique increment */
5980 size = (cnt * 68 / 25) % (sizeof(rb_string) - 1);
5982 len = size + sizeof(struct rb_item);
5984 started = rb_test_started;
5985 /* read rb_test_started before checking buffer enabled */
5988 event = ring_buffer_lock_reserve(data->buffer, len);
5990 /* Ignore dropped events before test starts. */
5993 data->bytes_dropped += len;
5995 data->bytes_dropped_nested += len;
6000 event_len = ring_buffer_event_length(event);
6002 if (RB_WARN_ON(data->buffer, event_len < len))
6005 item = ring_buffer_event_data(event);
6007 memcpy(item->str, rb_string, size);
6010 data->bytes_alloc_nested += event_len;
6011 data->bytes_written_nested += len;
6012 data->events_nested++;
6013 if (!data->min_size_nested || len < data->min_size_nested)
6014 data->min_size_nested = len;
6015 if (len > data->max_size_nested)
6016 data->max_size_nested = len;
6018 data->bytes_alloc += event_len;
6019 data->bytes_written += len;
6021 if (!data->min_size || len < data->min_size)
6022 data->max_size = len;
6023 if (len > data->max_size)
6024 data->max_size = len;
6028 ring_buffer_unlock_commit(data->buffer);
6033 static __init int rb_test(void *arg)
6035 struct rb_test_data *data = arg;
6037 while (!kthread_should_stop()) {
6038 rb_write_something(data, false);
6041 set_current_state(TASK_INTERRUPTIBLE);
6042 /* Now sleep between a min of 100-300us and a max of 1ms */
6043 usleep_range(((data->cnt % 3) + 1) * 100, 1000);
6049 static __init void rb_ipi(void *ignore)
6051 struct rb_test_data *data;
6052 int cpu = smp_processor_id();
6054 data = &rb_data[cpu];
6055 rb_write_something(data, true);
6058 static __init int rb_hammer_test(void *arg)
6060 while (!kthread_should_stop()) {
6062 /* Send an IPI to all cpus to write data! */
6063 smp_call_function(rb_ipi, NULL, 1);
6064 /* No sleep, but for non preempt, let others run */
6071 static __init int test_ringbuffer(void)
6073 struct task_struct *rb_hammer;
6074 struct trace_buffer *buffer;
6078 if (security_locked_down(LOCKDOWN_TRACEFS)) {
6079 pr_warn("Lockdown is enabled, skipping ring buffer tests\n");
6083 pr_info("Running ring buffer tests...\n");
6085 buffer = ring_buffer_alloc(RB_TEST_BUFFER_SIZE, RB_FL_OVERWRITE);
6086 if (WARN_ON(!buffer))
6089 /* Disable buffer so that threads can't write to it yet */
6090 ring_buffer_record_off(buffer);
6092 for_each_online_cpu(cpu) {
6093 rb_data[cpu].buffer = buffer;
6094 rb_data[cpu].cpu = cpu;
6095 rb_data[cpu].cnt = cpu;
6096 rb_threads[cpu] = kthread_run_on_cpu(rb_test, &rb_data[cpu],
6097 cpu, "rbtester/%u");
6098 if (WARN_ON(IS_ERR(rb_threads[cpu]))) {
6099 pr_cont("FAILED\n");
6100 ret = PTR_ERR(rb_threads[cpu]);
6105 /* Now create the rb hammer! */
6106 rb_hammer = kthread_run(rb_hammer_test, NULL, "rbhammer");
6107 if (WARN_ON(IS_ERR(rb_hammer))) {
6108 pr_cont("FAILED\n");
6109 ret = PTR_ERR(rb_hammer);
6113 ring_buffer_record_on(buffer);
6115 * Show buffer is enabled before setting rb_test_started.
6116 * Yes there's a small race window where events could be
6117 * dropped and the thread wont catch it. But when a ring
6118 * buffer gets enabled, there will always be some kind of
6119 * delay before other CPUs see it. Thus, we don't care about
6120 * those dropped events. We care about events dropped after
6121 * the threads see that the buffer is active.
6124 rb_test_started = true;
6126 set_current_state(TASK_INTERRUPTIBLE);
6127 /* Just run for 10 seconds */;
6128 schedule_timeout(10 * HZ);
6130 kthread_stop(rb_hammer);
6133 for_each_online_cpu(cpu) {
6134 if (!rb_threads[cpu])
6136 kthread_stop(rb_threads[cpu]);
6139 ring_buffer_free(buffer);
6144 pr_info("finished\n");
6145 for_each_online_cpu(cpu) {
6146 struct ring_buffer_event *event;
6147 struct rb_test_data *data = &rb_data[cpu];
6148 struct rb_item *item;
6149 unsigned long total_events;
6150 unsigned long total_dropped;
6151 unsigned long total_written;
6152 unsigned long total_alloc;
6153 unsigned long total_read = 0;
6154 unsigned long total_size = 0;
6155 unsigned long total_len = 0;
6156 unsigned long total_lost = 0;
6159 int small_event_size;
6163 total_events = data->events + data->events_nested;
6164 total_written = data->bytes_written + data->bytes_written_nested;
6165 total_alloc = data->bytes_alloc + data->bytes_alloc_nested;
6166 total_dropped = data->bytes_dropped + data->bytes_dropped_nested;
6168 big_event_size = data->max_size + data->max_size_nested;
6169 small_event_size = data->min_size + data->min_size_nested;
6171 pr_info("CPU %d:\n", cpu);
6172 pr_info(" events: %ld\n", total_events);
6173 pr_info(" dropped bytes: %ld\n", total_dropped);
6174 pr_info(" alloced bytes: %ld\n", total_alloc);
6175 pr_info(" written bytes: %ld\n", total_written);
6176 pr_info(" biggest event: %d\n", big_event_size);
6177 pr_info(" smallest event: %d\n", small_event_size);
6179 if (RB_WARN_ON(buffer, total_dropped))
6184 while ((event = ring_buffer_consume(buffer, cpu, NULL, &lost))) {
6186 item = ring_buffer_event_data(event);
6187 total_len += ring_buffer_event_length(event);
6188 total_size += item->size + sizeof(struct rb_item);
6189 if (memcmp(&item->str[0], rb_string, item->size) != 0) {
6190 pr_info("FAILED!\n");
6191 pr_info("buffer had: %.*s\n", item->size, item->str);
6192 pr_info("expected: %.*s\n", item->size, rb_string);
6193 RB_WARN_ON(buffer, 1);
6204 pr_info(" read events: %ld\n", total_read);
6205 pr_info(" lost events: %ld\n", total_lost);
6206 pr_info(" total events: %ld\n", total_lost + total_read);
6207 pr_info(" recorded len bytes: %ld\n", total_len);
6208 pr_info(" recorded size bytes: %ld\n", total_size);
6210 pr_info(" With dropped events, record len and size may not match\n"
6211 " alloced and written from above\n");
6213 if (RB_WARN_ON(buffer, total_len != total_alloc ||
6214 total_size != total_written))
6217 if (RB_WARN_ON(buffer, total_lost + total_read != total_events))
6223 pr_info("Ring buffer PASSED!\n");
6225 ring_buffer_free(buffer);
6229 late_initcall(test_ringbuffer);
6230 #endif /* CONFIG_RING_BUFFER_STARTUP_TEST */