4 * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
6 #include <linux/ring_buffer.h>
7 #include <linux/trace_clock.h>
8 #include <linux/spinlock.h>
9 #include <linux/debugfs.h>
10 #include <linux/uaccess.h>
11 #include <linux/hardirq.h>
12 #include <linux/kmemcheck.h>
13 #include <linux/module.h>
14 #include <linux/percpu.h>
15 #include <linux/mutex.h>
16 #include <linux/slab.h>
17 #include <linux/init.h>
18 #include <linux/hash.h>
19 #include <linux/list.h>
20 #include <linux/cpu.h>
23 #include <asm/local.h>
26 static void update_pages_handler(struct work_struct *work);
29 * The ring buffer header is special. We must manually up keep it.
31 int ring_buffer_print_entry_header(struct trace_seq *s)
35 ret = trace_seq_printf(s, "# compressed entry header\n");
36 ret = trace_seq_printf(s, "\ttype_len : 5 bits\n");
37 ret = trace_seq_printf(s, "\ttime_delta : 27 bits\n");
38 ret = trace_seq_printf(s, "\tarray : 32 bits\n");
39 ret = trace_seq_printf(s, "\n");
40 ret = trace_seq_printf(s, "\tpadding : type == %d\n",
41 RINGBUF_TYPE_PADDING);
42 ret = trace_seq_printf(s, "\ttime_extend : type == %d\n",
43 RINGBUF_TYPE_TIME_EXTEND);
44 ret = trace_seq_printf(s, "\tdata max type_len == %d\n",
45 RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
51 * The ring buffer is made up of a list of pages. A separate list of pages is
52 * allocated for each CPU. A writer may only write to a buffer that is
53 * associated with the CPU it is currently executing on. A reader may read
54 * from any per cpu buffer.
56 * The reader is special. For each per cpu buffer, the reader has its own
57 * reader page. When a reader has read the entire reader page, this reader
58 * page is swapped with another page in the ring buffer.
60 * Now, as long as the writer is off the reader page, the reader can do what
61 * ever it wants with that page. The writer will never write to that page
62 * again (as long as it is out of the ring buffer).
64 * Here's some silly ASCII art.
67 * |reader| RING BUFFER
69 * +------+ +---+ +---+ +---+
78 * |reader| RING BUFFER
79 * |page |------------------v
80 * +------+ +---+ +---+ +---+
89 * |reader| RING BUFFER
90 * |page |------------------v
91 * +------+ +---+ +---+ +---+
96 * +------------------------------+
100 * |buffer| RING BUFFER
101 * |page |------------------v
102 * +------+ +---+ +---+ +---+
104 * | New +---+ +---+ +---+
107 * +------------------------------+
110 * After we make this swap, the reader can hand this page off to the splice
111 * code and be done with it. It can even allocate a new page if it needs to
112 * and swap that into the ring buffer.
114 * We will be using cmpxchg soon to make all this lockless.
119 * A fast way to enable or disable all ring buffers is to
120 * call tracing_on or tracing_off. Turning off the ring buffers
121 * prevents all ring buffers from being recorded to.
122 * Turning this switch on, makes it OK to write to the
123 * ring buffer, if the ring buffer is enabled itself.
125 * There's three layers that must be on in order to write
126 * to the ring buffer.
128 * 1) This global flag must be set.
129 * 2) The ring buffer must be enabled for recording.
130 * 3) The per cpu buffer must be enabled for recording.
132 * In case of an anomaly, this global flag has a bit set that
133 * will permantly disable all ring buffers.
137 * Global flag to disable all recording to ring buffers
138 * This has two bits: ON, DISABLED
142 * 0 0 : ring buffers are off
143 * 1 0 : ring buffers are on
144 * X 1 : ring buffers are permanently disabled
148 RB_BUFFERS_ON_BIT = 0,
149 RB_BUFFERS_DISABLED_BIT = 1,
153 RB_BUFFERS_ON = 1 << RB_BUFFERS_ON_BIT,
154 RB_BUFFERS_DISABLED = 1 << RB_BUFFERS_DISABLED_BIT,
157 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
159 /* Used for individual buffers (after the counter) */
160 #define RB_BUFFER_OFF (1 << 20)
162 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
165 * tracing_off_permanent - permanently disable ring buffers
167 * This function, once called, will disable all ring buffers
170 void tracing_off_permanent(void)
172 set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
175 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
176 #define RB_ALIGNMENT 4U
177 #define RB_MAX_SMALL_DATA (RB_ALIGNMENT * RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
178 #define RB_EVNT_MIN_SIZE 8U /* two 32bit words */
180 #if !defined(CONFIG_64BIT) || defined(CONFIG_HAVE_EFFICIENT_UNALIGNED_ACCESS)
181 # define RB_FORCE_8BYTE_ALIGNMENT 0
182 # define RB_ARCH_ALIGNMENT RB_ALIGNMENT
184 # define RB_FORCE_8BYTE_ALIGNMENT 1
185 # define RB_ARCH_ALIGNMENT 8U
188 /* define RINGBUF_TYPE_DATA for 'case RINGBUF_TYPE_DATA:' */
189 #define RINGBUF_TYPE_DATA 0 ... RINGBUF_TYPE_DATA_TYPE_LEN_MAX
192 RB_LEN_TIME_EXTEND = 8,
193 RB_LEN_TIME_STAMP = 16,
196 #define skip_time_extend(event) \
197 ((struct ring_buffer_event *)((char *)event + RB_LEN_TIME_EXTEND))
199 static inline int rb_null_event(struct ring_buffer_event *event)
201 return event->type_len == RINGBUF_TYPE_PADDING && !event->time_delta;
204 static void rb_event_set_padding(struct ring_buffer_event *event)
206 /* padding has a NULL time_delta */
207 event->type_len = RINGBUF_TYPE_PADDING;
208 event->time_delta = 0;
212 rb_event_data_length(struct ring_buffer_event *event)
217 length = event->type_len * RB_ALIGNMENT;
219 length = event->array[0];
220 return length + RB_EVNT_HDR_SIZE;
224 * Return the length of the given event. Will return
225 * the length of the time extend if the event is a
228 static inline unsigned
229 rb_event_length(struct ring_buffer_event *event)
231 switch (event->type_len) {
232 case RINGBUF_TYPE_PADDING:
233 if (rb_null_event(event))
236 return event->array[0] + RB_EVNT_HDR_SIZE;
238 case RINGBUF_TYPE_TIME_EXTEND:
239 return RB_LEN_TIME_EXTEND;
241 case RINGBUF_TYPE_TIME_STAMP:
242 return RB_LEN_TIME_STAMP;
244 case RINGBUF_TYPE_DATA:
245 return rb_event_data_length(event);
254 * Return total length of time extend and data,
255 * or just the event length for all other events.
257 static inline unsigned
258 rb_event_ts_length(struct ring_buffer_event *event)
262 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
263 /* time extends include the data event after it */
264 len = RB_LEN_TIME_EXTEND;
265 event = skip_time_extend(event);
267 return len + rb_event_length(event);
271 * ring_buffer_event_length - return the length of the event
272 * @event: the event to get the length of
274 * Returns the size of the data load of a data event.
275 * If the event is something other than a data event, it
276 * returns the size of the event itself. With the exception
277 * of a TIME EXTEND, where it still returns the size of the
278 * data load of the data event after it.
280 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
284 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
285 event = skip_time_extend(event);
287 length = rb_event_length(event);
288 if (event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
290 length -= RB_EVNT_HDR_SIZE;
291 if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
292 length -= sizeof(event->array[0]);
295 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
297 /* inline for ring buffer fast paths */
299 rb_event_data(struct ring_buffer_event *event)
301 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
302 event = skip_time_extend(event);
303 BUG_ON(event->type_len > RINGBUF_TYPE_DATA_TYPE_LEN_MAX);
304 /* If length is in len field, then array[0] has the data */
306 return (void *)&event->array[0];
307 /* Otherwise length is in array[0] and array[1] has the data */
308 return (void *)&event->array[1];
312 * ring_buffer_event_data - return the data of the event
313 * @event: the event to get the data from
315 void *ring_buffer_event_data(struct ring_buffer_event *event)
317 return rb_event_data(event);
319 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
321 #define for_each_buffer_cpu(buffer, cpu) \
322 for_each_cpu(cpu, buffer->cpumask)
325 #define TS_MASK ((1ULL << TS_SHIFT) - 1)
326 #define TS_DELTA_TEST (~TS_MASK)
328 /* Flag when events were overwritten */
329 #define RB_MISSED_EVENTS (1 << 31)
330 /* Missed count stored at end */
331 #define RB_MISSED_STORED (1 << 30)
333 struct buffer_data_page {
334 u64 time_stamp; /* page time stamp */
335 local_t commit; /* write committed index */
336 unsigned char data[]; /* data of buffer page */
340 * Note, the buffer_page list must be first. The buffer pages
341 * are allocated in cache lines, which means that each buffer
342 * page will be at the beginning of a cache line, and thus
343 * the least significant bits will be zero. We use this to
344 * add flags in the list struct pointers, to make the ring buffer
348 struct list_head list; /* list of buffer pages */
349 local_t write; /* index for next write */
350 unsigned read; /* index for next read */
351 local_t entries; /* entries on this page */
352 unsigned long real_end; /* real end of data */
353 struct buffer_data_page *page; /* Actual data page */
357 * The buffer page counters, write and entries, must be reset
358 * atomically when crossing page boundaries. To synchronize this
359 * update, two counters are inserted into the number. One is
360 * the actual counter for the write position or count on the page.
362 * The other is a counter of updaters. Before an update happens
363 * the update partition of the counter is incremented. This will
364 * allow the updater to update the counter atomically.
366 * The counter is 20 bits, and the state data is 12.
368 #define RB_WRITE_MASK 0xfffff
369 #define RB_WRITE_INTCNT (1 << 20)
371 static void rb_init_page(struct buffer_data_page *bpage)
373 local_set(&bpage->commit, 0);
377 * ring_buffer_page_len - the size of data on the page.
378 * @page: The page to read
380 * Returns the amount of data on the page, including buffer page header.
382 size_t ring_buffer_page_len(void *page)
384 return local_read(&((struct buffer_data_page *)page)->commit)
389 * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
392 static void free_buffer_page(struct buffer_page *bpage)
394 free_page((unsigned long)bpage->page);
399 * We need to fit the time_stamp delta into 27 bits.
401 static inline int test_time_stamp(u64 delta)
403 if (delta & TS_DELTA_TEST)
408 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
410 /* Max payload is BUF_PAGE_SIZE - header (8bytes) */
411 #define BUF_MAX_DATA_SIZE (BUF_PAGE_SIZE - (sizeof(u32) * 2))
413 int ring_buffer_print_page_header(struct trace_seq *s)
415 struct buffer_data_page field;
418 ret = trace_seq_printf(s, "\tfield: u64 timestamp;\t"
419 "offset:0;\tsize:%u;\tsigned:%u;\n",
420 (unsigned int)sizeof(field.time_stamp),
421 (unsigned int)is_signed_type(u64));
423 ret = trace_seq_printf(s, "\tfield: local_t commit;\t"
424 "offset:%u;\tsize:%u;\tsigned:%u;\n",
425 (unsigned int)offsetof(typeof(field), commit),
426 (unsigned int)sizeof(field.commit),
427 (unsigned int)is_signed_type(long));
429 ret = trace_seq_printf(s, "\tfield: int overwrite;\t"
430 "offset:%u;\tsize:%u;\tsigned:%u;\n",
431 (unsigned int)offsetof(typeof(field), commit),
433 (unsigned int)is_signed_type(long));
435 ret = trace_seq_printf(s, "\tfield: char data;\t"
436 "offset:%u;\tsize:%u;\tsigned:%u;\n",
437 (unsigned int)offsetof(typeof(field), data),
438 (unsigned int)BUF_PAGE_SIZE,
439 (unsigned int)is_signed_type(char));
445 * head_page == tail_page && head == tail then buffer is empty.
447 struct ring_buffer_per_cpu {
449 atomic_t record_disabled;
450 struct ring_buffer *buffer;
451 raw_spinlock_t reader_lock; /* serialize readers */
452 arch_spinlock_t lock;
453 struct lock_class_key lock_key;
454 unsigned int nr_pages;
455 struct list_head *pages;
456 struct buffer_page *head_page; /* read from head */
457 struct buffer_page *tail_page; /* write to tail */
458 struct buffer_page *commit_page; /* committed pages */
459 struct buffer_page *reader_page;
460 unsigned long lost_events;
461 unsigned long last_overrun;
462 local_t entries_bytes;
463 local_t commit_overrun;
469 unsigned long read_bytes;
472 /* ring buffer pages to update, > 0 to add, < 0 to remove */
473 int nr_pages_to_update;
474 struct list_head new_pages; /* new pages to add */
475 struct work_struct update_pages_work;
476 struct completion update_done;
482 atomic_t record_disabled;
483 atomic_t resize_disabled;
484 cpumask_var_t cpumask;
486 struct lock_class_key *reader_lock_key;
490 struct ring_buffer_per_cpu **buffers;
492 #ifdef CONFIG_HOTPLUG_CPU
493 struct notifier_block cpu_notify;
498 struct ring_buffer_iter {
499 struct ring_buffer_per_cpu *cpu_buffer;
501 struct buffer_page *head_page;
502 struct buffer_page *cache_reader_page;
503 unsigned long cache_read;
507 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
508 #define RB_WARN_ON(b, cond) \
510 int _____ret = unlikely(cond); \
512 if (__same_type(*(b), struct ring_buffer_per_cpu)) { \
513 struct ring_buffer_per_cpu *__b = \
515 atomic_inc(&__b->buffer->record_disabled); \
517 atomic_inc(&b->record_disabled); \
523 /* Up this if you want to test the TIME_EXTENTS and normalization */
524 #define DEBUG_SHIFT 0
526 static inline u64 rb_time_stamp(struct ring_buffer *buffer)
528 /* shift to debug/test normalization and TIME_EXTENTS */
529 return buffer->clock() << DEBUG_SHIFT;
532 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
536 preempt_disable_notrace();
537 time = rb_time_stamp(buffer);
538 preempt_enable_no_resched_notrace();
542 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
544 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
547 /* Just stupid testing the normalize function and deltas */
550 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
553 * Making the ring buffer lockless makes things tricky.
554 * Although writes only happen on the CPU that they are on,
555 * and they only need to worry about interrupts. Reads can
558 * The reader page is always off the ring buffer, but when the
559 * reader finishes with a page, it needs to swap its page with
560 * a new one from the buffer. The reader needs to take from
561 * the head (writes go to the tail). But if a writer is in overwrite
562 * mode and wraps, it must push the head page forward.
564 * Here lies the problem.
566 * The reader must be careful to replace only the head page, and
567 * not another one. As described at the top of the file in the
568 * ASCII art, the reader sets its old page to point to the next
569 * page after head. It then sets the page after head to point to
570 * the old reader page. But if the writer moves the head page
571 * during this operation, the reader could end up with the tail.
573 * We use cmpxchg to help prevent this race. We also do something
574 * special with the page before head. We set the LSB to 1.
576 * When the writer must push the page forward, it will clear the
577 * bit that points to the head page, move the head, and then set
578 * the bit that points to the new head page.
580 * We also don't want an interrupt coming in and moving the head
581 * page on another writer. Thus we use the second LSB to catch
584 * head->list->prev->next bit 1 bit 0
587 * Points to head page 0 1
590 * Note we can not trust the prev pointer of the head page, because:
592 * +----+ +-----+ +-----+
593 * | |------>| T |---X--->| N |
595 * +----+ +-----+ +-----+
598 * +----------| R |----------+ |
602 * Key: ---X--> HEAD flag set in pointer
607 * (see __rb_reserve_next() to see where this happens)
609 * What the above shows is that the reader just swapped out
610 * the reader page with a page in the buffer, but before it
611 * could make the new header point back to the new page added
612 * it was preempted by a writer. The writer moved forward onto
613 * the new page added by the reader and is about to move forward
616 * You can see, it is legitimate for the previous pointer of
617 * the head (or any page) not to point back to itself. But only
621 #define RB_PAGE_NORMAL 0UL
622 #define RB_PAGE_HEAD 1UL
623 #define RB_PAGE_UPDATE 2UL
626 #define RB_FLAG_MASK 3UL
628 /* PAGE_MOVED is not part of the mask */
629 #define RB_PAGE_MOVED 4UL
632 * rb_list_head - remove any bit
634 static struct list_head *rb_list_head(struct list_head *list)
636 unsigned long val = (unsigned long)list;
638 return (struct list_head *)(val & ~RB_FLAG_MASK);
642 * rb_is_head_page - test if the given page is the head page
644 * Because the reader may move the head_page pointer, we can
645 * not trust what the head page is (it may be pointing to
646 * the reader page). But if the next page is a header page,
647 * its flags will be non zero.
650 rb_is_head_page(struct ring_buffer_per_cpu *cpu_buffer,
651 struct buffer_page *page, struct list_head *list)
655 val = (unsigned long)list->next;
657 if ((val & ~RB_FLAG_MASK) != (unsigned long)&page->list)
658 return RB_PAGE_MOVED;
660 return val & RB_FLAG_MASK;
666 * The unique thing about the reader page, is that, if the
667 * writer is ever on it, the previous pointer never points
668 * back to the reader page.
670 static int rb_is_reader_page(struct buffer_page *page)
672 struct list_head *list = page->list.prev;
674 return rb_list_head(list->next) != &page->list;
678 * rb_set_list_to_head - set a list_head to be pointing to head.
680 static void rb_set_list_to_head(struct ring_buffer_per_cpu *cpu_buffer,
681 struct list_head *list)
685 ptr = (unsigned long *)&list->next;
686 *ptr |= RB_PAGE_HEAD;
687 *ptr &= ~RB_PAGE_UPDATE;
691 * rb_head_page_activate - sets up head page
693 static void rb_head_page_activate(struct ring_buffer_per_cpu *cpu_buffer)
695 struct buffer_page *head;
697 head = cpu_buffer->head_page;
702 * Set the previous list pointer to have the HEAD flag.
704 rb_set_list_to_head(cpu_buffer, head->list.prev);
707 static void rb_list_head_clear(struct list_head *list)
709 unsigned long *ptr = (unsigned long *)&list->next;
711 *ptr &= ~RB_FLAG_MASK;
715 * rb_head_page_dactivate - clears head page ptr (for free list)
718 rb_head_page_deactivate(struct ring_buffer_per_cpu *cpu_buffer)
720 struct list_head *hd;
722 /* Go through the whole list and clear any pointers found. */
723 rb_list_head_clear(cpu_buffer->pages);
725 list_for_each(hd, cpu_buffer->pages)
726 rb_list_head_clear(hd);
729 static int rb_head_page_set(struct ring_buffer_per_cpu *cpu_buffer,
730 struct buffer_page *head,
731 struct buffer_page *prev,
732 int old_flag, int new_flag)
734 struct list_head *list;
735 unsigned long val = (unsigned long)&head->list;
740 val &= ~RB_FLAG_MASK;
742 ret = cmpxchg((unsigned long *)&list->next,
743 val | old_flag, val | new_flag);
745 /* check if the reader took the page */
746 if ((ret & ~RB_FLAG_MASK) != val)
747 return RB_PAGE_MOVED;
749 return ret & RB_FLAG_MASK;
752 static int rb_head_page_set_update(struct ring_buffer_per_cpu *cpu_buffer,
753 struct buffer_page *head,
754 struct buffer_page *prev,
757 return rb_head_page_set(cpu_buffer, head, prev,
758 old_flag, RB_PAGE_UPDATE);
761 static int rb_head_page_set_head(struct ring_buffer_per_cpu *cpu_buffer,
762 struct buffer_page *head,
763 struct buffer_page *prev,
766 return rb_head_page_set(cpu_buffer, head, prev,
767 old_flag, RB_PAGE_HEAD);
770 static int rb_head_page_set_normal(struct ring_buffer_per_cpu *cpu_buffer,
771 struct buffer_page *head,
772 struct buffer_page *prev,
775 return rb_head_page_set(cpu_buffer, head, prev,
776 old_flag, RB_PAGE_NORMAL);
779 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
780 struct buffer_page **bpage)
782 struct list_head *p = rb_list_head((*bpage)->list.next);
784 *bpage = list_entry(p, struct buffer_page, list);
787 static struct buffer_page *
788 rb_set_head_page(struct ring_buffer_per_cpu *cpu_buffer)
790 struct buffer_page *head;
791 struct buffer_page *page;
792 struct list_head *list;
795 if (RB_WARN_ON(cpu_buffer, !cpu_buffer->head_page))
799 list = cpu_buffer->pages;
800 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev->next) != list))
803 page = head = cpu_buffer->head_page;
805 * It is possible that the writer moves the header behind
806 * where we started, and we miss in one loop.
807 * A second loop should grab the header, but we'll do
808 * three loops just because I'm paranoid.
810 for (i = 0; i < 3; i++) {
812 if (rb_is_head_page(cpu_buffer, page, page->list.prev)) {
813 cpu_buffer->head_page = page;
816 rb_inc_page(cpu_buffer, &page);
817 } while (page != head);
820 RB_WARN_ON(cpu_buffer, 1);
825 static int rb_head_page_replace(struct buffer_page *old,
826 struct buffer_page *new)
828 unsigned long *ptr = (unsigned long *)&old->list.prev->next;
832 val = *ptr & ~RB_FLAG_MASK;
835 ret = cmpxchg(ptr, val, (unsigned long)&new->list);
841 * rb_tail_page_update - move the tail page forward
843 * Returns 1 if moved tail page, 0 if someone else did.
845 static int rb_tail_page_update(struct ring_buffer_per_cpu *cpu_buffer,
846 struct buffer_page *tail_page,
847 struct buffer_page *next_page)
849 struct buffer_page *old_tail;
850 unsigned long old_entries;
851 unsigned long old_write;
855 * The tail page now needs to be moved forward.
857 * We need to reset the tail page, but without messing
858 * with possible erasing of data brought in by interrupts
859 * that have moved the tail page and are currently on it.
861 * We add a counter to the write field to denote this.
863 old_write = local_add_return(RB_WRITE_INTCNT, &next_page->write);
864 old_entries = local_add_return(RB_WRITE_INTCNT, &next_page->entries);
867 * Just make sure we have seen our old_write and synchronize
868 * with any interrupts that come in.
873 * If the tail page is still the same as what we think
874 * it is, then it is up to us to update the tail
877 if (tail_page == cpu_buffer->tail_page) {
878 /* Zero the write counter */
879 unsigned long val = old_write & ~RB_WRITE_MASK;
880 unsigned long eval = old_entries & ~RB_WRITE_MASK;
883 * This will only succeed if an interrupt did
884 * not come in and change it. In which case, we
885 * do not want to modify it.
887 * We add (void) to let the compiler know that we do not care
888 * about the return value of these functions. We use the
889 * cmpxchg to only update if an interrupt did not already
890 * do it for us. If the cmpxchg fails, we don't care.
892 (void)local_cmpxchg(&next_page->write, old_write, val);
893 (void)local_cmpxchg(&next_page->entries, old_entries, eval);
896 * No need to worry about races with clearing out the commit.
897 * it only can increment when a commit takes place. But that
898 * only happens in the outer most nested commit.
900 local_set(&next_page->page->commit, 0);
902 old_tail = cmpxchg(&cpu_buffer->tail_page,
903 tail_page, next_page);
905 if (old_tail == tail_page)
912 static int rb_check_bpage(struct ring_buffer_per_cpu *cpu_buffer,
913 struct buffer_page *bpage)
915 unsigned long val = (unsigned long)bpage;
917 if (RB_WARN_ON(cpu_buffer, val & RB_FLAG_MASK))
924 * rb_check_list - make sure a pointer to a list has the last bits zero
926 static int rb_check_list(struct ring_buffer_per_cpu *cpu_buffer,
927 struct list_head *list)
929 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->prev) != list->prev))
931 if (RB_WARN_ON(cpu_buffer, rb_list_head(list->next) != list->next))
937 * check_pages - integrity check of buffer pages
938 * @cpu_buffer: CPU buffer with pages to test
940 * As a safety measure we check to make sure the data pages have not
943 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
945 struct list_head *head = cpu_buffer->pages;
946 struct buffer_page *bpage, *tmp;
948 /* Reset the head page if it exists */
949 if (cpu_buffer->head_page)
950 rb_set_head_page(cpu_buffer);
952 rb_head_page_deactivate(cpu_buffer);
954 if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
956 if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
959 if (rb_check_list(cpu_buffer, head))
962 list_for_each_entry_safe(bpage, tmp, head, list) {
963 if (RB_WARN_ON(cpu_buffer,
964 bpage->list.next->prev != &bpage->list))
966 if (RB_WARN_ON(cpu_buffer,
967 bpage->list.prev->next != &bpage->list))
969 if (rb_check_list(cpu_buffer, &bpage->list))
973 rb_head_page_activate(cpu_buffer);
978 static int __rb_allocate_pages(int nr_pages, struct list_head *pages, int cpu)
981 struct buffer_page *bpage, *tmp;
983 for (i = 0; i < nr_pages; i++) {
986 * __GFP_NORETRY flag makes sure that the allocation fails
987 * gracefully without invoking oom-killer and the system is
990 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
991 GFP_KERNEL | __GFP_NORETRY,
996 list_add(&bpage->list, pages);
998 page = alloc_pages_node(cpu_to_node(cpu),
999 GFP_KERNEL | __GFP_NORETRY, 0);
1002 bpage->page = page_address(page);
1003 rb_init_page(bpage->page);
1009 list_for_each_entry_safe(bpage, tmp, pages, list) {
1010 list_del_init(&bpage->list);
1011 free_buffer_page(bpage);
1017 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
1024 if (__rb_allocate_pages(nr_pages, &pages, cpu_buffer->cpu))
1028 * The ring buffer page list is a circular list that does not
1029 * start and end with a list head. All page list items point to
1032 cpu_buffer->pages = pages.next;
1035 cpu_buffer->nr_pages = nr_pages;
1037 rb_check_pages(cpu_buffer);
1042 static struct ring_buffer_per_cpu *
1043 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int nr_pages, int cpu)
1045 struct ring_buffer_per_cpu *cpu_buffer;
1046 struct buffer_page *bpage;
1050 cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
1051 GFP_KERNEL, cpu_to_node(cpu));
1055 cpu_buffer->cpu = cpu;
1056 cpu_buffer->buffer = buffer;
1057 raw_spin_lock_init(&cpu_buffer->reader_lock);
1058 lockdep_set_class(&cpu_buffer->reader_lock, buffer->reader_lock_key);
1059 cpu_buffer->lock = (arch_spinlock_t)__ARCH_SPIN_LOCK_UNLOCKED;
1060 INIT_WORK(&cpu_buffer->update_pages_work, update_pages_handler);
1061 init_completion(&cpu_buffer->update_done);
1063 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
1064 GFP_KERNEL, cpu_to_node(cpu));
1066 goto fail_free_buffer;
1068 rb_check_bpage(cpu_buffer, bpage);
1070 cpu_buffer->reader_page = bpage;
1071 page = alloc_pages_node(cpu_to_node(cpu), GFP_KERNEL, 0);
1073 goto fail_free_reader;
1074 bpage->page = page_address(page);
1075 rb_init_page(bpage->page);
1077 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
1079 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1081 goto fail_free_reader;
1083 cpu_buffer->head_page
1084 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1085 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1087 rb_head_page_activate(cpu_buffer);
1092 free_buffer_page(cpu_buffer->reader_page);
1099 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1101 struct list_head *head = cpu_buffer->pages;
1102 struct buffer_page *bpage, *tmp;
1104 free_buffer_page(cpu_buffer->reader_page);
1106 rb_head_page_deactivate(cpu_buffer);
1109 list_for_each_entry_safe(bpage, tmp, head, list) {
1110 list_del_init(&bpage->list);
1111 free_buffer_page(bpage);
1113 bpage = list_entry(head, struct buffer_page, list);
1114 free_buffer_page(bpage);
1120 #ifdef CONFIG_HOTPLUG_CPU
1121 static int rb_cpu_notify(struct notifier_block *self,
1122 unsigned long action, void *hcpu);
1126 * ring_buffer_alloc - allocate a new ring_buffer
1127 * @size: the size in bytes per cpu that is needed.
1128 * @flags: attributes to set for the ring buffer.
1130 * Currently the only flag that is available is the RB_FL_OVERWRITE
1131 * flag. This flag means that the buffer will overwrite old data
1132 * when the buffer wraps. If this flag is not set, the buffer will
1133 * drop data when the tail hits the head.
1135 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1136 struct lock_class_key *key)
1138 struct ring_buffer *buffer;
1142 /* keep it in its own cache line */
1143 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1148 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1149 goto fail_free_buffer;
1151 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1152 buffer->flags = flags;
1153 buffer->clock = trace_clock_local;
1154 buffer->reader_lock_key = key;
1156 /* need at least two pages */
1161 * In case of non-hotplug cpu, if the ring-buffer is allocated
1162 * in early initcall, it will not be notified of secondary cpus.
1163 * In that off case, we need to allocate for all possible cpus.
1165 #ifdef CONFIG_HOTPLUG_CPU
1167 cpumask_copy(buffer->cpumask, cpu_online_mask);
1169 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1171 buffer->cpus = nr_cpu_ids;
1173 bsize = sizeof(void *) * nr_cpu_ids;
1174 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1176 if (!buffer->buffers)
1177 goto fail_free_cpumask;
1179 for_each_buffer_cpu(buffer, cpu) {
1180 buffer->buffers[cpu] =
1181 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1182 if (!buffer->buffers[cpu])
1183 goto fail_free_buffers;
1186 #ifdef CONFIG_HOTPLUG_CPU
1187 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1188 buffer->cpu_notify.priority = 0;
1189 register_cpu_notifier(&buffer->cpu_notify);
1193 mutex_init(&buffer->mutex);
1198 for_each_buffer_cpu(buffer, cpu) {
1199 if (buffer->buffers[cpu])
1200 rb_free_cpu_buffer(buffer->buffers[cpu]);
1202 kfree(buffer->buffers);
1205 free_cpumask_var(buffer->cpumask);
1212 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1215 * ring_buffer_free - free a ring buffer.
1216 * @buffer: the buffer to free.
1219 ring_buffer_free(struct ring_buffer *buffer)
1225 #ifdef CONFIG_HOTPLUG_CPU
1226 unregister_cpu_notifier(&buffer->cpu_notify);
1229 for_each_buffer_cpu(buffer, cpu)
1230 rb_free_cpu_buffer(buffer->buffers[cpu]);
1234 kfree(buffer->buffers);
1235 free_cpumask_var(buffer->cpumask);
1239 EXPORT_SYMBOL_GPL(ring_buffer_free);
1241 void ring_buffer_set_clock(struct ring_buffer *buffer,
1244 buffer->clock = clock;
1247 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1249 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1251 return local_read(&bpage->entries) & RB_WRITE_MASK;
1254 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1256 return local_read(&bpage->write) & RB_WRITE_MASK;
1260 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned int nr_pages)
1262 struct list_head *tail_page, *to_remove, *next_page;
1263 struct buffer_page *to_remove_page, *tmp_iter_page;
1264 struct buffer_page *last_page, *first_page;
1265 unsigned int nr_removed;
1266 unsigned long head_bit;
1271 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1272 atomic_inc(&cpu_buffer->record_disabled);
1274 * We don't race with the readers since we have acquired the reader
1275 * lock. We also don't race with writers after disabling recording.
1276 * This makes it easy to figure out the first and the last page to be
1277 * removed from the list. We unlink all the pages in between including
1278 * the first and last pages. This is done in a busy loop so that we
1279 * lose the least number of traces.
1280 * The pages are freed after we restart recording and unlock readers.
1282 tail_page = &cpu_buffer->tail_page->list;
1285 * tail page might be on reader page, we remove the next page
1286 * from the ring buffer
1288 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1289 tail_page = rb_list_head(tail_page->next);
1290 to_remove = tail_page;
1292 /* start of pages to remove */
1293 first_page = list_entry(rb_list_head(to_remove->next),
1294 struct buffer_page, list);
1296 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1297 to_remove = rb_list_head(to_remove)->next;
1298 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1301 next_page = rb_list_head(to_remove)->next;
1304 * Now we remove all pages between tail_page and next_page.
1305 * Make sure that we have head_bit value preserved for the
1308 tail_page->next = (struct list_head *)((unsigned long)next_page |
1310 next_page = rb_list_head(next_page);
1311 next_page->prev = tail_page;
1313 /* make sure pages points to a valid page in the ring buffer */
1314 cpu_buffer->pages = next_page;
1316 /* update head page */
1318 cpu_buffer->head_page = list_entry(next_page,
1319 struct buffer_page, list);
1322 * change read pointer to make sure any read iterators reset
1325 cpu_buffer->read = 0;
1327 /* pages are removed, resume tracing and then free the pages */
1328 atomic_dec(&cpu_buffer->record_disabled);
1329 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1331 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1333 /* last buffer page to remove */
1334 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1336 tmp_iter_page = first_page;
1339 to_remove_page = tmp_iter_page;
1340 rb_inc_page(cpu_buffer, &tmp_iter_page);
1342 /* update the counters */
1343 page_entries = rb_page_entries(to_remove_page);
1346 * If something was added to this page, it was full
1347 * since it is not the tail page. So we deduct the
1348 * bytes consumed in ring buffer from here.
1349 * No need to update overruns, since this page is
1350 * deleted from ring buffer and its entries are
1351 * already accounted for.
1353 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1357 * We have already removed references to this list item, just
1358 * free up the buffer_page and its page
1360 free_buffer_page(to_remove_page);
1363 } while (to_remove_page != last_page);
1365 RB_WARN_ON(cpu_buffer, nr_removed);
1367 return nr_removed == 0;
1371 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1373 struct list_head *pages = &cpu_buffer->new_pages;
1374 int retries, success;
1376 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1378 * We are holding the reader lock, so the reader page won't be swapped
1379 * in the ring buffer. Now we are racing with the writer trying to
1380 * move head page and the tail page.
1381 * We are going to adapt the reader page update process where:
1382 * 1. We first splice the start and end of list of new pages between
1383 * the head page and its previous page.
1384 * 2. We cmpxchg the prev_page->next to point from head page to the
1385 * start of new pages list.
1386 * 3. Finally, we update the head->prev to the end of new list.
1388 * We will try this process 10 times, to make sure that we don't keep
1394 struct list_head *head_page, *prev_page, *r;
1395 struct list_head *last_page, *first_page;
1396 struct list_head *head_page_with_bit;
1398 head_page = &rb_set_head_page(cpu_buffer)->list;
1399 prev_page = head_page->prev;
1401 first_page = pages->next;
1402 last_page = pages->prev;
1404 head_page_with_bit = (struct list_head *)
1405 ((unsigned long)head_page | RB_PAGE_HEAD);
1407 last_page->next = head_page_with_bit;
1408 first_page->prev = prev_page;
1410 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1412 if (r == head_page_with_bit) {
1414 * yay, we replaced the page pointer to our new list,
1415 * now, we just have to update to head page's prev
1416 * pointer to point to end of list
1418 head_page->prev = last_page;
1425 INIT_LIST_HEAD(pages);
1427 * If we weren't successful in adding in new pages, warn and stop
1430 RB_WARN_ON(cpu_buffer, !success);
1431 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1433 /* free pages if they weren't inserted */
1435 struct buffer_page *bpage, *tmp;
1436 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1438 list_del_init(&bpage->list);
1439 free_buffer_page(bpage);
1445 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1449 if (cpu_buffer->nr_pages_to_update > 0)
1450 success = rb_insert_pages(cpu_buffer);
1452 success = rb_remove_pages(cpu_buffer,
1453 -cpu_buffer->nr_pages_to_update);
1456 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1459 static void update_pages_handler(struct work_struct *work)
1461 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1462 struct ring_buffer_per_cpu, update_pages_work);
1463 rb_update_pages(cpu_buffer);
1464 complete(&cpu_buffer->update_done);
1468 * ring_buffer_resize - resize the ring buffer
1469 * @buffer: the buffer to resize.
1470 * @size: the new size.
1472 * Minimum size is 2 * BUF_PAGE_SIZE.
1474 * Returns 0 on success and < 0 on failure.
1476 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1479 struct ring_buffer_per_cpu *cpu_buffer;
1484 * Always succeed at resizing a non-existent buffer:
1489 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1490 size *= BUF_PAGE_SIZE;
1492 /* we need a minimum of two pages */
1493 if (size < BUF_PAGE_SIZE * 2)
1494 size = BUF_PAGE_SIZE * 2;
1496 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1499 * Don't succeed if resizing is disabled, as a reader might be
1500 * manipulating the ring buffer and is expecting a sane state while
1503 if (atomic_read(&buffer->resize_disabled))
1506 /* prevent another thread from changing buffer sizes */
1507 mutex_lock(&buffer->mutex);
1509 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1510 /* calculate the pages to update */
1511 for_each_buffer_cpu(buffer, cpu) {
1512 cpu_buffer = buffer->buffers[cpu];
1514 cpu_buffer->nr_pages_to_update = nr_pages -
1515 cpu_buffer->nr_pages;
1517 * nothing more to do for removing pages or no update
1519 if (cpu_buffer->nr_pages_to_update <= 0)
1522 * to add pages, make sure all new pages can be
1523 * allocated without receiving ENOMEM
1525 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1526 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1527 &cpu_buffer->new_pages, cpu)) {
1528 /* not enough memory for new pages */
1536 * Fire off all the required work handlers
1537 * We can't schedule on offline CPUs, but it's not necessary
1538 * since we can change their buffer sizes without any race.
1540 for_each_buffer_cpu(buffer, cpu) {
1541 cpu_buffer = buffer->buffers[cpu];
1542 if (!cpu_buffer->nr_pages_to_update)
1545 if (cpu_online(cpu))
1546 schedule_work_on(cpu,
1547 &cpu_buffer->update_pages_work);
1549 rb_update_pages(cpu_buffer);
1552 /* wait for all the updates to complete */
1553 for_each_buffer_cpu(buffer, cpu) {
1554 cpu_buffer = buffer->buffers[cpu];
1555 if (!cpu_buffer->nr_pages_to_update)
1558 if (cpu_online(cpu))
1559 wait_for_completion(&cpu_buffer->update_done);
1560 cpu_buffer->nr_pages_to_update = 0;
1565 cpu_buffer = buffer->buffers[cpu_id];
1567 if (nr_pages == cpu_buffer->nr_pages)
1570 cpu_buffer->nr_pages_to_update = nr_pages -
1571 cpu_buffer->nr_pages;
1573 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1574 if (cpu_buffer->nr_pages_to_update > 0 &&
1575 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1576 &cpu_buffer->new_pages, cpu_id)) {
1583 if (cpu_online(cpu_id)) {
1584 schedule_work_on(cpu_id,
1585 &cpu_buffer->update_pages_work);
1586 wait_for_completion(&cpu_buffer->update_done);
1588 rb_update_pages(cpu_buffer);
1590 cpu_buffer->nr_pages_to_update = 0;
1596 * The ring buffer resize can happen with the ring buffer
1597 * enabled, so that the update disturbs the tracing as little
1598 * as possible. But if the buffer is disabled, we do not need
1599 * to worry about that, and we can take the time to verify
1600 * that the buffer is not corrupt.
1602 if (atomic_read(&buffer->record_disabled)) {
1603 atomic_inc(&buffer->record_disabled);
1605 * Even though the buffer was disabled, we must make sure
1606 * that it is truly disabled before calling rb_check_pages.
1607 * There could have been a race between checking
1608 * record_disable and incrementing it.
1610 synchronize_sched();
1611 for_each_buffer_cpu(buffer, cpu) {
1612 cpu_buffer = buffer->buffers[cpu];
1613 rb_check_pages(cpu_buffer);
1615 atomic_dec(&buffer->record_disabled);
1618 mutex_unlock(&buffer->mutex);
1622 for_each_buffer_cpu(buffer, cpu) {
1623 struct buffer_page *bpage, *tmp;
1625 cpu_buffer = buffer->buffers[cpu];
1626 cpu_buffer->nr_pages_to_update = 0;
1628 if (list_empty(&cpu_buffer->new_pages))
1631 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1633 list_del_init(&bpage->list);
1634 free_buffer_page(bpage);
1637 mutex_unlock(&buffer->mutex);
1640 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1642 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1644 mutex_lock(&buffer->mutex);
1646 buffer->flags |= RB_FL_OVERWRITE;
1648 buffer->flags &= ~RB_FL_OVERWRITE;
1649 mutex_unlock(&buffer->mutex);
1651 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1653 static inline void *
1654 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1656 return bpage->data + index;
1659 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1661 return bpage->page->data + index;
1664 static inline struct ring_buffer_event *
1665 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1667 return __rb_page_index(cpu_buffer->reader_page,
1668 cpu_buffer->reader_page->read);
1671 static inline struct ring_buffer_event *
1672 rb_iter_head_event(struct ring_buffer_iter *iter)
1674 return __rb_page_index(iter->head_page, iter->head);
1677 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1679 return local_read(&bpage->page->commit);
1682 /* Size is determined by what has been committed */
1683 static inline unsigned rb_page_size(struct buffer_page *bpage)
1685 return rb_page_commit(bpage);
1688 static inline unsigned
1689 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1691 return rb_page_commit(cpu_buffer->commit_page);
1694 static inline unsigned
1695 rb_event_index(struct ring_buffer_event *event)
1697 unsigned long addr = (unsigned long)event;
1699 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1703 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1704 struct ring_buffer_event *event)
1706 unsigned long addr = (unsigned long)event;
1707 unsigned long index;
1709 index = rb_event_index(event);
1712 return cpu_buffer->commit_page->page == (void *)addr &&
1713 rb_commit_index(cpu_buffer) == index;
1717 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1719 unsigned long max_count;
1722 * We only race with interrupts and NMIs on this CPU.
1723 * If we own the commit event, then we can commit
1724 * all others that interrupted us, since the interruptions
1725 * are in stack format (they finish before they come
1726 * back to us). This allows us to do a simple loop to
1727 * assign the commit to the tail.
1730 max_count = cpu_buffer->nr_pages * 100;
1732 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1733 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1735 if (RB_WARN_ON(cpu_buffer,
1736 rb_is_reader_page(cpu_buffer->tail_page)))
1738 local_set(&cpu_buffer->commit_page->page->commit,
1739 rb_page_write(cpu_buffer->commit_page));
1740 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1741 cpu_buffer->write_stamp =
1742 cpu_buffer->commit_page->page->time_stamp;
1743 /* add barrier to keep gcc from optimizing too much */
1746 while (rb_commit_index(cpu_buffer) !=
1747 rb_page_write(cpu_buffer->commit_page)) {
1749 local_set(&cpu_buffer->commit_page->page->commit,
1750 rb_page_write(cpu_buffer->commit_page));
1751 RB_WARN_ON(cpu_buffer,
1752 local_read(&cpu_buffer->commit_page->page->commit) &
1757 /* again, keep gcc from optimizing */
1761 * If an interrupt came in just after the first while loop
1762 * and pushed the tail page forward, we will be left with
1763 * a dangling commit that will never go forward.
1765 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1769 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1771 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1772 cpu_buffer->reader_page->read = 0;
1775 static void rb_inc_iter(struct ring_buffer_iter *iter)
1777 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1780 * The iterator could be on the reader page (it starts there).
1781 * But the head could have moved, since the reader was
1782 * found. Check for this case and assign the iterator
1783 * to the head page instead of next.
1785 if (iter->head_page == cpu_buffer->reader_page)
1786 iter->head_page = rb_set_head_page(cpu_buffer);
1788 rb_inc_page(cpu_buffer, &iter->head_page);
1790 iter->read_stamp = iter->head_page->page->time_stamp;
1794 /* Slow path, do not inline */
1795 static noinline struct ring_buffer_event *
1796 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1798 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1800 /* Not the first event on the page? */
1801 if (rb_event_index(event)) {
1802 event->time_delta = delta & TS_MASK;
1803 event->array[0] = delta >> TS_SHIFT;
1805 /* nope, just zero it */
1806 event->time_delta = 0;
1807 event->array[0] = 0;
1810 return skip_time_extend(event);
1814 * ring_buffer_update_event - update event type and data
1815 * @event: the even to update
1816 * @type: the type of event
1817 * @length: the size of the event field in the ring buffer
1819 * Update the type and data fields of the event. The length
1820 * is the actual size that is written to the ring buffer,
1821 * and with this, we can determine what to place into the
1825 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1826 struct ring_buffer_event *event, unsigned length,
1827 int add_timestamp, u64 delta)
1829 /* Only a commit updates the timestamp */
1830 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
1834 * If we need to add a timestamp, then we
1835 * add it to the start of the resevered space.
1837 if (unlikely(add_timestamp)) {
1838 event = rb_add_time_stamp(event, delta);
1839 length -= RB_LEN_TIME_EXTEND;
1843 event->time_delta = delta;
1844 length -= RB_EVNT_HDR_SIZE;
1845 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
1846 event->type_len = 0;
1847 event->array[0] = length;
1849 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1853 * rb_handle_head_page - writer hit the head page
1855 * Returns: +1 to retry page
1860 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1861 struct buffer_page *tail_page,
1862 struct buffer_page *next_page)
1864 struct buffer_page *new_head;
1869 entries = rb_page_entries(next_page);
1872 * The hard part is here. We need to move the head
1873 * forward, and protect against both readers on
1874 * other CPUs and writers coming in via interrupts.
1876 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1880 * type can be one of four:
1881 * NORMAL - an interrupt already moved it for us
1882 * HEAD - we are the first to get here.
1883 * UPDATE - we are the interrupt interrupting
1885 * MOVED - a reader on another CPU moved the next
1886 * pointer to its reader page. Give up
1893 * We changed the head to UPDATE, thus
1894 * it is our responsibility to update
1897 local_add(entries, &cpu_buffer->overrun);
1898 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1901 * The entries will be zeroed out when we move the
1905 /* still more to do */
1908 case RB_PAGE_UPDATE:
1910 * This is an interrupt that interrupt the
1911 * previous update. Still more to do.
1914 case RB_PAGE_NORMAL:
1916 * An interrupt came in before the update
1917 * and processed this for us.
1918 * Nothing left to do.
1923 * The reader is on another CPU and just did
1924 * a swap with our next_page.
1929 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1934 * Now that we are here, the old head pointer is
1935 * set to UPDATE. This will keep the reader from
1936 * swapping the head page with the reader page.
1937 * The reader (on another CPU) will spin till
1940 * We just need to protect against interrupts
1941 * doing the job. We will set the next pointer
1942 * to HEAD. After that, we set the old pointer
1943 * to NORMAL, but only if it was HEAD before.
1944 * otherwise we are an interrupt, and only
1945 * want the outer most commit to reset it.
1947 new_head = next_page;
1948 rb_inc_page(cpu_buffer, &new_head);
1950 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1954 * Valid returns are:
1955 * HEAD - an interrupt came in and already set it.
1956 * NORMAL - One of two things:
1957 * 1) We really set it.
1958 * 2) A bunch of interrupts came in and moved
1959 * the page forward again.
1963 case RB_PAGE_NORMAL:
1967 RB_WARN_ON(cpu_buffer, 1);
1972 * It is possible that an interrupt came in,
1973 * set the head up, then more interrupts came in
1974 * and moved it again. When we get back here,
1975 * the page would have been set to NORMAL but we
1976 * just set it back to HEAD.
1978 * How do you detect this? Well, if that happened
1979 * the tail page would have moved.
1981 if (ret == RB_PAGE_NORMAL) {
1983 * If the tail had moved passed next, then we need
1984 * to reset the pointer.
1986 if (cpu_buffer->tail_page != tail_page &&
1987 cpu_buffer->tail_page != next_page)
1988 rb_head_page_set_normal(cpu_buffer, new_head,
1994 * If this was the outer most commit (the one that
1995 * changed the original pointer from HEAD to UPDATE),
1996 * then it is up to us to reset it to NORMAL.
1998 if (type == RB_PAGE_HEAD) {
1999 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2002 if (RB_WARN_ON(cpu_buffer,
2003 ret != RB_PAGE_UPDATE))
2010 static unsigned rb_calculate_event_length(unsigned length)
2012 struct ring_buffer_event event; /* Used only for sizeof array */
2014 /* zero length can cause confusions */
2018 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2019 length += sizeof(event.array[0]);
2021 length += RB_EVNT_HDR_SIZE;
2022 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2028 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2029 struct buffer_page *tail_page,
2030 unsigned long tail, unsigned long length)
2032 struct ring_buffer_event *event;
2035 * Only the event that crossed the page boundary
2036 * must fill the old tail_page with padding.
2038 if (tail >= BUF_PAGE_SIZE) {
2040 * If the page was filled, then we still need
2041 * to update the real_end. Reset it to zero
2042 * and the reader will ignore it.
2044 if (tail == BUF_PAGE_SIZE)
2045 tail_page->real_end = 0;
2047 local_sub(length, &tail_page->write);
2051 event = __rb_page_index(tail_page, tail);
2052 kmemcheck_annotate_bitfield(event, bitfield);
2054 /* account for padding bytes */
2055 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2058 * Save the original length to the meta data.
2059 * This will be used by the reader to add lost event
2062 tail_page->real_end = tail;
2065 * If this event is bigger than the minimum size, then
2066 * we need to be careful that we don't subtract the
2067 * write counter enough to allow another writer to slip
2069 * We put in a discarded commit instead, to make sure
2070 * that this space is not used again.
2072 * If we are less than the minimum size, we don't need to
2075 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2076 /* No room for any events */
2078 /* Mark the rest of the page with padding */
2079 rb_event_set_padding(event);
2081 /* Set the write back to the previous setting */
2082 local_sub(length, &tail_page->write);
2086 /* Put in a discarded event */
2087 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2088 event->type_len = RINGBUF_TYPE_PADDING;
2089 /* time delta must be non zero */
2090 event->time_delta = 1;
2092 /* Set write to end of buffer */
2093 length = (tail + length) - BUF_PAGE_SIZE;
2094 local_sub(length, &tail_page->write);
2098 * This is the slow path, force gcc not to inline it.
2100 static noinline struct ring_buffer_event *
2101 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2102 unsigned long length, unsigned long tail,
2103 struct buffer_page *tail_page, u64 ts)
2105 struct buffer_page *commit_page = cpu_buffer->commit_page;
2106 struct ring_buffer *buffer = cpu_buffer->buffer;
2107 struct buffer_page *next_page;
2110 next_page = tail_page;
2112 rb_inc_page(cpu_buffer, &next_page);
2115 * If for some reason, we had an interrupt storm that made
2116 * it all the way around the buffer, bail, and warn
2119 if (unlikely(next_page == commit_page)) {
2120 local_inc(&cpu_buffer->commit_overrun);
2125 * This is where the fun begins!
2127 * We are fighting against races between a reader that
2128 * could be on another CPU trying to swap its reader
2129 * page with the buffer head.
2131 * We are also fighting against interrupts coming in and
2132 * moving the head or tail on us as well.
2134 * If the next page is the head page then we have filled
2135 * the buffer, unless the commit page is still on the
2138 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2141 * If the commit is not on the reader page, then
2142 * move the header page.
2144 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2146 * If we are not in overwrite mode,
2147 * this is easy, just stop here.
2149 if (!(buffer->flags & RB_FL_OVERWRITE))
2152 ret = rb_handle_head_page(cpu_buffer,
2161 * We need to be careful here too. The
2162 * commit page could still be on the reader
2163 * page. We could have a small buffer, and
2164 * have filled up the buffer with events
2165 * from interrupts and such, and wrapped.
2167 * Note, if the tail page is also the on the
2168 * reader_page, we let it move out.
2170 if (unlikely((cpu_buffer->commit_page !=
2171 cpu_buffer->tail_page) &&
2172 (cpu_buffer->commit_page ==
2173 cpu_buffer->reader_page))) {
2174 local_inc(&cpu_buffer->commit_overrun);
2180 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2183 * Nested commits always have zero deltas, so
2184 * just reread the time stamp
2186 ts = rb_time_stamp(buffer);
2187 next_page->page->time_stamp = ts;
2192 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2194 /* fail and let the caller try again */
2195 return ERR_PTR(-EAGAIN);
2199 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2204 static struct ring_buffer_event *
2205 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2206 unsigned long length, u64 ts,
2207 u64 delta, int add_timestamp)
2209 struct buffer_page *tail_page;
2210 struct ring_buffer_event *event;
2211 unsigned long tail, write;
2214 * If the time delta since the last event is too big to
2215 * hold in the time field of the event, then we append a
2216 * TIME EXTEND event ahead of the data event.
2218 if (unlikely(add_timestamp))
2219 length += RB_LEN_TIME_EXTEND;
2221 tail_page = cpu_buffer->tail_page;
2222 write = local_add_return(length, &tail_page->write);
2224 /* set write to only the index of the write */
2225 write &= RB_WRITE_MASK;
2226 tail = write - length;
2228 /* See if we shot pass the end of this buffer page */
2229 if (unlikely(write > BUF_PAGE_SIZE))
2230 return rb_move_tail(cpu_buffer, length, tail,
2233 /* We reserved something on the buffer */
2235 event = __rb_page_index(tail_page, tail);
2236 kmemcheck_annotate_bitfield(event, bitfield);
2237 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2239 local_inc(&tail_page->entries);
2242 * If this is the first commit on the page, then update
2246 tail_page->page->time_stamp = ts;
2248 /* account for these added bytes */
2249 local_add(length, &cpu_buffer->entries_bytes);
2255 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2256 struct ring_buffer_event *event)
2258 unsigned long new_index, old_index;
2259 struct buffer_page *bpage;
2260 unsigned long index;
2263 new_index = rb_event_index(event);
2264 old_index = new_index + rb_event_ts_length(event);
2265 addr = (unsigned long)event;
2268 bpage = cpu_buffer->tail_page;
2270 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2271 unsigned long write_mask =
2272 local_read(&bpage->write) & ~RB_WRITE_MASK;
2273 unsigned long event_length = rb_event_length(event);
2275 * This is on the tail page. It is possible that
2276 * a write could come in and move the tail page
2277 * and write to the next page. That is fine
2278 * because we just shorten what is on this page.
2280 old_index += write_mask;
2281 new_index += write_mask;
2282 index = local_cmpxchg(&bpage->write, old_index, new_index);
2283 if (index == old_index) {
2284 /* update counters */
2285 local_sub(event_length, &cpu_buffer->entries_bytes);
2290 /* could not discard */
2294 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2296 local_inc(&cpu_buffer->committing);
2297 local_inc(&cpu_buffer->commits);
2300 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2302 unsigned long commits;
2304 if (RB_WARN_ON(cpu_buffer,
2305 !local_read(&cpu_buffer->committing)))
2309 commits = local_read(&cpu_buffer->commits);
2310 /* synchronize with interrupts */
2312 if (local_read(&cpu_buffer->committing) == 1)
2313 rb_set_commit_to_write(cpu_buffer);
2315 local_dec(&cpu_buffer->committing);
2317 /* synchronize with interrupts */
2321 * Need to account for interrupts coming in between the
2322 * updating of the commit page and the clearing of the
2323 * committing counter.
2325 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2326 !local_read(&cpu_buffer->committing)) {
2327 local_inc(&cpu_buffer->committing);
2332 static struct ring_buffer_event *
2333 rb_reserve_next_event(struct ring_buffer *buffer,
2334 struct ring_buffer_per_cpu *cpu_buffer,
2335 unsigned long length)
2337 struct ring_buffer_event *event;
2343 rb_start_commit(cpu_buffer);
2345 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2347 * Due to the ability to swap a cpu buffer from a buffer
2348 * it is possible it was swapped before we committed.
2349 * (committing stops a swap). We check for it here and
2350 * if it happened, we have to fail the write.
2353 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2354 local_dec(&cpu_buffer->committing);
2355 local_dec(&cpu_buffer->commits);
2360 length = rb_calculate_event_length(length);
2366 * We allow for interrupts to reenter here and do a trace.
2367 * If one does, it will cause this original code to loop
2368 * back here. Even with heavy interrupts happening, this
2369 * should only happen a few times in a row. If this happens
2370 * 1000 times in a row, there must be either an interrupt
2371 * storm or we have something buggy.
2374 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2377 ts = rb_time_stamp(cpu_buffer->buffer);
2378 diff = ts - cpu_buffer->write_stamp;
2380 /* make sure this diff is calculated here */
2383 /* Did the write stamp get updated already? */
2384 if (likely(ts >= cpu_buffer->write_stamp)) {
2386 if (unlikely(test_time_stamp(delta))) {
2387 int local_clock_stable = 1;
2388 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2389 local_clock_stable = sched_clock_stable;
2391 WARN_ONCE(delta > (1ULL << 59),
2392 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2393 (unsigned long long)delta,
2394 (unsigned long long)ts,
2395 (unsigned long long)cpu_buffer->write_stamp,
2396 local_clock_stable ? "" :
2397 "If you just came from a suspend/resume,\n"
2398 "please switch to the trace global clock:\n"
2399 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2404 event = __rb_reserve_next(cpu_buffer, length, ts,
2405 delta, add_timestamp);
2406 if (unlikely(PTR_ERR(event) == -EAGAIN))
2415 rb_end_commit(cpu_buffer);
2419 #ifdef CONFIG_TRACING
2421 #define TRACE_RECURSIVE_DEPTH 16
2423 /* Keep this code out of the fast path cache */
2424 static noinline void trace_recursive_fail(void)
2426 /* Disable all tracing before we do anything else */
2427 tracing_off_permanent();
2429 printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2430 "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2431 trace_recursion_buffer(),
2432 hardirq_count() >> HARDIRQ_SHIFT,
2433 softirq_count() >> SOFTIRQ_SHIFT,
2439 static inline int trace_recursive_lock(void)
2441 trace_recursion_inc();
2443 if (likely(trace_recursion_buffer() < TRACE_RECURSIVE_DEPTH))
2446 trace_recursive_fail();
2451 static inline void trace_recursive_unlock(void)
2453 WARN_ON_ONCE(!trace_recursion_buffer());
2455 trace_recursion_dec();
2460 #define trace_recursive_lock() (0)
2461 #define trace_recursive_unlock() do { } while (0)
2466 * ring_buffer_lock_reserve - reserve a part of the buffer
2467 * @buffer: the ring buffer to reserve from
2468 * @length: the length of the data to reserve (excluding event header)
2470 * Returns a reseverd event on the ring buffer to copy directly to.
2471 * The user of this interface will need to get the body to write into
2472 * and can use the ring_buffer_event_data() interface.
2474 * The length is the length of the data needed, not the event length
2475 * which also includes the event header.
2477 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2478 * If NULL is returned, then nothing has been allocated or locked.
2480 struct ring_buffer_event *
2481 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2483 struct ring_buffer_per_cpu *cpu_buffer;
2484 struct ring_buffer_event *event;
2487 if (ring_buffer_flags != RB_BUFFERS_ON)
2490 /* If we are tracing schedule, we don't want to recurse */
2491 preempt_disable_notrace();
2493 if (atomic_read(&buffer->record_disabled))
2496 if (trace_recursive_lock())
2499 cpu = raw_smp_processor_id();
2501 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2504 cpu_buffer = buffer->buffers[cpu];
2506 if (atomic_read(&cpu_buffer->record_disabled))
2509 if (length > BUF_MAX_DATA_SIZE)
2512 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2519 trace_recursive_unlock();
2522 preempt_enable_notrace();
2525 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2528 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2529 struct ring_buffer_event *event)
2534 * The event first in the commit queue updates the
2537 if (rb_event_is_commit(cpu_buffer, event)) {
2539 * A commit event that is first on a page
2540 * updates the write timestamp with the page stamp
2542 if (!rb_event_index(event))
2543 cpu_buffer->write_stamp =
2544 cpu_buffer->commit_page->page->time_stamp;
2545 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2546 delta = event->array[0];
2548 delta += event->time_delta;
2549 cpu_buffer->write_stamp += delta;
2551 cpu_buffer->write_stamp += event->time_delta;
2555 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2556 struct ring_buffer_event *event)
2558 local_inc(&cpu_buffer->entries);
2559 rb_update_write_stamp(cpu_buffer, event);
2560 rb_end_commit(cpu_buffer);
2564 * ring_buffer_unlock_commit - commit a reserved
2565 * @buffer: The buffer to commit to
2566 * @event: The event pointer to commit.
2568 * This commits the data to the ring buffer, and releases any locks held.
2570 * Must be paired with ring_buffer_lock_reserve.
2572 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2573 struct ring_buffer_event *event)
2575 struct ring_buffer_per_cpu *cpu_buffer;
2576 int cpu = raw_smp_processor_id();
2578 cpu_buffer = buffer->buffers[cpu];
2580 rb_commit(cpu_buffer, event);
2582 trace_recursive_unlock();
2584 preempt_enable_notrace();
2588 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2590 static inline void rb_event_discard(struct ring_buffer_event *event)
2592 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2593 event = skip_time_extend(event);
2595 /* array[0] holds the actual length for the discarded event */
2596 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2597 event->type_len = RINGBUF_TYPE_PADDING;
2598 /* time delta must be non zero */
2599 if (!event->time_delta)
2600 event->time_delta = 1;
2604 * Decrement the entries to the page that an event is on.
2605 * The event does not even need to exist, only the pointer
2606 * to the page it is on. This may only be called before the commit
2610 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2611 struct ring_buffer_event *event)
2613 unsigned long addr = (unsigned long)event;
2614 struct buffer_page *bpage = cpu_buffer->commit_page;
2615 struct buffer_page *start;
2619 /* Do the likely case first */
2620 if (likely(bpage->page == (void *)addr)) {
2621 local_dec(&bpage->entries);
2626 * Because the commit page may be on the reader page we
2627 * start with the next page and check the end loop there.
2629 rb_inc_page(cpu_buffer, &bpage);
2632 if (bpage->page == (void *)addr) {
2633 local_dec(&bpage->entries);
2636 rb_inc_page(cpu_buffer, &bpage);
2637 } while (bpage != start);
2639 /* commit not part of this buffer?? */
2640 RB_WARN_ON(cpu_buffer, 1);
2644 * ring_buffer_commit_discard - discard an event that has not been committed
2645 * @buffer: the ring buffer
2646 * @event: non committed event to discard
2648 * Sometimes an event that is in the ring buffer needs to be ignored.
2649 * This function lets the user discard an event in the ring buffer
2650 * and then that event will not be read later.
2652 * This function only works if it is called before the the item has been
2653 * committed. It will try to free the event from the ring buffer
2654 * if another event has not been added behind it.
2656 * If another event has been added behind it, it will set the event
2657 * up as discarded, and perform the commit.
2659 * If this function is called, do not call ring_buffer_unlock_commit on
2662 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2663 struct ring_buffer_event *event)
2665 struct ring_buffer_per_cpu *cpu_buffer;
2668 /* The event is discarded regardless */
2669 rb_event_discard(event);
2671 cpu = smp_processor_id();
2672 cpu_buffer = buffer->buffers[cpu];
2675 * This must only be called if the event has not been
2676 * committed yet. Thus we can assume that preemption
2677 * is still disabled.
2679 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2681 rb_decrement_entry(cpu_buffer, event);
2682 if (rb_try_to_discard(cpu_buffer, event))
2686 * The commit is still visible by the reader, so we
2687 * must still update the timestamp.
2689 rb_update_write_stamp(cpu_buffer, event);
2691 rb_end_commit(cpu_buffer);
2693 trace_recursive_unlock();
2695 preempt_enable_notrace();
2698 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2701 * ring_buffer_write - write data to the buffer without reserving
2702 * @buffer: The ring buffer to write to.
2703 * @length: The length of the data being written (excluding the event header)
2704 * @data: The data to write to the buffer.
2706 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2707 * one function. If you already have the data to write to the buffer, it
2708 * may be easier to simply call this function.
2710 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2711 * and not the length of the event which would hold the header.
2713 int ring_buffer_write(struct ring_buffer *buffer,
2714 unsigned long length,
2717 struct ring_buffer_per_cpu *cpu_buffer;
2718 struct ring_buffer_event *event;
2723 if (ring_buffer_flags != RB_BUFFERS_ON)
2726 preempt_disable_notrace();
2728 if (atomic_read(&buffer->record_disabled))
2731 cpu = raw_smp_processor_id();
2733 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2736 cpu_buffer = buffer->buffers[cpu];
2738 if (atomic_read(&cpu_buffer->record_disabled))
2741 if (length > BUF_MAX_DATA_SIZE)
2744 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2748 body = rb_event_data(event);
2750 memcpy(body, data, length);
2752 rb_commit(cpu_buffer, event);
2756 preempt_enable_notrace();
2760 EXPORT_SYMBOL_GPL(ring_buffer_write);
2762 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2764 struct buffer_page *reader = cpu_buffer->reader_page;
2765 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2766 struct buffer_page *commit = cpu_buffer->commit_page;
2768 /* In case of error, head will be NULL */
2769 if (unlikely(!head))
2772 return reader->read == rb_page_commit(reader) &&
2773 (commit == reader ||
2775 head->read == rb_page_commit(commit)));
2779 * ring_buffer_record_disable - stop all writes into the buffer
2780 * @buffer: The ring buffer to stop writes to.
2782 * This prevents all writes to the buffer. Any attempt to write
2783 * to the buffer after this will fail and return NULL.
2785 * The caller should call synchronize_sched() after this.
2787 void ring_buffer_record_disable(struct ring_buffer *buffer)
2789 atomic_inc(&buffer->record_disabled);
2791 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2794 * ring_buffer_record_enable - enable writes to the buffer
2795 * @buffer: The ring buffer to enable writes
2797 * Note, multiple disables will need the same number of enables
2798 * to truly enable the writing (much like preempt_disable).
2800 void ring_buffer_record_enable(struct ring_buffer *buffer)
2802 atomic_dec(&buffer->record_disabled);
2804 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2807 * ring_buffer_record_off - stop all writes into the buffer
2808 * @buffer: The ring buffer to stop writes to.
2810 * This prevents all writes to the buffer. Any attempt to write
2811 * to the buffer after this will fail and return NULL.
2813 * This is different than ring_buffer_record_disable() as
2814 * it works like an on/off switch, where as the disable() verison
2815 * must be paired with a enable().
2817 void ring_buffer_record_off(struct ring_buffer *buffer)
2820 unsigned int new_rd;
2823 rd = atomic_read(&buffer->record_disabled);
2824 new_rd = rd | RB_BUFFER_OFF;
2825 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
2827 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
2830 * ring_buffer_record_on - restart writes into the buffer
2831 * @buffer: The ring buffer to start writes to.
2833 * This enables all writes to the buffer that was disabled by
2834 * ring_buffer_record_off().
2836 * This is different than ring_buffer_record_enable() as
2837 * it works like an on/off switch, where as the enable() verison
2838 * must be paired with a disable().
2840 void ring_buffer_record_on(struct ring_buffer *buffer)
2843 unsigned int new_rd;
2846 rd = atomic_read(&buffer->record_disabled);
2847 new_rd = rd & ~RB_BUFFER_OFF;
2848 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
2850 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
2853 * ring_buffer_record_is_on - return true if the ring buffer can write
2854 * @buffer: The ring buffer to see if write is enabled
2856 * Returns true if the ring buffer is in a state that it accepts writes.
2858 int ring_buffer_record_is_on(struct ring_buffer *buffer)
2860 return !atomic_read(&buffer->record_disabled);
2864 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2865 * @buffer: The ring buffer to stop writes to.
2866 * @cpu: The CPU buffer to stop
2868 * This prevents all writes to the buffer. Any attempt to write
2869 * to the buffer after this will fail and return NULL.
2871 * The caller should call synchronize_sched() after this.
2873 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2875 struct ring_buffer_per_cpu *cpu_buffer;
2877 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2880 cpu_buffer = buffer->buffers[cpu];
2881 atomic_inc(&cpu_buffer->record_disabled);
2883 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2886 * ring_buffer_record_enable_cpu - enable writes to the buffer
2887 * @buffer: The ring buffer to enable writes
2888 * @cpu: The CPU to enable.
2890 * Note, multiple disables will need the same number of enables
2891 * to truly enable the writing (much like preempt_disable).
2893 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2895 struct ring_buffer_per_cpu *cpu_buffer;
2897 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2900 cpu_buffer = buffer->buffers[cpu];
2901 atomic_dec(&cpu_buffer->record_disabled);
2903 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2906 * The total entries in the ring buffer is the running counter
2907 * of entries entered into the ring buffer, minus the sum of
2908 * the entries read from the ring buffer and the number of
2909 * entries that were overwritten.
2911 static inline unsigned long
2912 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
2914 return local_read(&cpu_buffer->entries) -
2915 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
2919 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
2920 * @buffer: The ring buffer
2921 * @cpu: The per CPU buffer to read from.
2923 unsigned long ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
2925 unsigned long flags;
2926 struct ring_buffer_per_cpu *cpu_buffer;
2927 struct buffer_page *bpage;
2930 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2933 cpu_buffer = buffer->buffers[cpu];
2934 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2936 * if the tail is on reader_page, oldest time stamp is on the reader
2939 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
2940 bpage = cpu_buffer->reader_page;
2942 bpage = rb_set_head_page(cpu_buffer);
2943 ret = bpage->page->time_stamp;
2944 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2948 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
2951 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
2952 * @buffer: The ring buffer
2953 * @cpu: The per CPU buffer to read from.
2955 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
2957 struct ring_buffer_per_cpu *cpu_buffer;
2960 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2963 cpu_buffer = buffer->buffers[cpu];
2964 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
2968 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
2971 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2972 * @buffer: The ring buffer
2973 * @cpu: The per CPU buffer to get the entries from.
2975 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2977 struct ring_buffer_per_cpu *cpu_buffer;
2979 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2982 cpu_buffer = buffer->buffers[cpu];
2984 return rb_num_of_entries(cpu_buffer);
2986 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
2989 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
2990 * @buffer: The ring buffer
2991 * @cpu: The per CPU buffer to get the number of overruns from
2993 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
2995 struct ring_buffer_per_cpu *cpu_buffer;
2998 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3001 cpu_buffer = buffer->buffers[cpu];
3002 ret = local_read(&cpu_buffer->overrun);
3006 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3009 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
3010 * @buffer: The ring buffer
3011 * @cpu: The per CPU buffer to get the number of overruns from
3014 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3016 struct ring_buffer_per_cpu *cpu_buffer;
3019 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3022 cpu_buffer = buffer->buffers[cpu];
3023 ret = local_read(&cpu_buffer->commit_overrun);
3027 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3030 * ring_buffer_entries - get the number of entries in a buffer
3031 * @buffer: The ring buffer
3033 * Returns the total number of entries in the ring buffer
3036 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3038 struct ring_buffer_per_cpu *cpu_buffer;
3039 unsigned long entries = 0;
3042 /* if you care about this being correct, lock the buffer */
3043 for_each_buffer_cpu(buffer, cpu) {
3044 cpu_buffer = buffer->buffers[cpu];
3045 entries += rb_num_of_entries(cpu_buffer);
3050 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3053 * ring_buffer_overruns - get the number of overruns in buffer
3054 * @buffer: The ring buffer
3056 * Returns the total number of overruns in the ring buffer
3059 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3061 struct ring_buffer_per_cpu *cpu_buffer;
3062 unsigned long overruns = 0;
3065 /* if you care about this being correct, lock the buffer */
3066 for_each_buffer_cpu(buffer, cpu) {
3067 cpu_buffer = buffer->buffers[cpu];
3068 overruns += local_read(&cpu_buffer->overrun);
3073 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3075 static void rb_iter_reset(struct ring_buffer_iter *iter)
3077 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3079 /* Iterator usage is expected to have record disabled */
3080 if (list_empty(&cpu_buffer->reader_page->list)) {
3081 iter->head_page = rb_set_head_page(cpu_buffer);
3082 if (unlikely(!iter->head_page))
3084 iter->head = iter->head_page->read;
3086 iter->head_page = cpu_buffer->reader_page;
3087 iter->head = cpu_buffer->reader_page->read;
3090 iter->read_stamp = cpu_buffer->read_stamp;
3092 iter->read_stamp = iter->head_page->page->time_stamp;
3093 iter->cache_reader_page = cpu_buffer->reader_page;
3094 iter->cache_read = cpu_buffer->read;
3098 * ring_buffer_iter_reset - reset an iterator
3099 * @iter: The iterator to reset
3101 * Resets the iterator, so that it will start from the beginning
3104 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3106 struct ring_buffer_per_cpu *cpu_buffer;
3107 unsigned long flags;
3112 cpu_buffer = iter->cpu_buffer;
3114 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3115 rb_iter_reset(iter);
3116 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3118 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3121 * ring_buffer_iter_empty - check if an iterator has no more to read
3122 * @iter: The iterator to check
3124 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3126 struct ring_buffer_per_cpu *cpu_buffer;
3128 cpu_buffer = iter->cpu_buffer;
3130 return iter->head_page == cpu_buffer->commit_page &&
3131 iter->head == rb_commit_index(cpu_buffer);
3133 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3136 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3137 struct ring_buffer_event *event)
3141 switch (event->type_len) {
3142 case RINGBUF_TYPE_PADDING:
3145 case RINGBUF_TYPE_TIME_EXTEND:
3146 delta = event->array[0];
3148 delta += event->time_delta;
3149 cpu_buffer->read_stamp += delta;
3152 case RINGBUF_TYPE_TIME_STAMP:
3153 /* FIXME: not implemented */
3156 case RINGBUF_TYPE_DATA:
3157 cpu_buffer->read_stamp += event->time_delta;
3167 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3168 struct ring_buffer_event *event)
3172 switch (event->type_len) {
3173 case RINGBUF_TYPE_PADDING:
3176 case RINGBUF_TYPE_TIME_EXTEND:
3177 delta = event->array[0];
3179 delta += event->time_delta;
3180 iter->read_stamp += delta;
3183 case RINGBUF_TYPE_TIME_STAMP:
3184 /* FIXME: not implemented */
3187 case RINGBUF_TYPE_DATA:
3188 iter->read_stamp += event->time_delta;
3197 static struct buffer_page *
3198 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3200 struct buffer_page *reader = NULL;
3201 unsigned long overwrite;
3202 unsigned long flags;
3206 local_irq_save(flags);
3207 arch_spin_lock(&cpu_buffer->lock);
3211 * This should normally only loop twice. But because the
3212 * start of the reader inserts an empty page, it causes
3213 * a case where we will loop three times. There should be no
3214 * reason to loop four times (that I know of).
3216 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3221 reader = cpu_buffer->reader_page;
3223 /* If there's more to read, return this page */
3224 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3227 /* Never should we have an index greater than the size */
3228 if (RB_WARN_ON(cpu_buffer,
3229 cpu_buffer->reader_page->read > rb_page_size(reader)))
3232 /* check if we caught up to the tail */
3234 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3238 * Reset the reader page to size zero.
3240 local_set(&cpu_buffer->reader_page->write, 0);
3241 local_set(&cpu_buffer->reader_page->entries, 0);
3242 local_set(&cpu_buffer->reader_page->page->commit, 0);
3243 cpu_buffer->reader_page->real_end = 0;
3247 * Splice the empty reader page into the list around the head.
3249 reader = rb_set_head_page(cpu_buffer);
3250 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3251 cpu_buffer->reader_page->list.prev = reader->list.prev;
3254 * cpu_buffer->pages just needs to point to the buffer, it
3255 * has no specific buffer page to point to. Lets move it out
3256 * of our way so we don't accidentally swap it.
3258 cpu_buffer->pages = reader->list.prev;
3260 /* The reader page will be pointing to the new head */
3261 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3264 * We want to make sure we read the overruns after we set up our
3265 * pointers to the next object. The writer side does a
3266 * cmpxchg to cross pages which acts as the mb on the writer
3267 * side. Note, the reader will constantly fail the swap
3268 * while the writer is updating the pointers, so this
3269 * guarantees that the overwrite recorded here is the one we
3270 * want to compare with the last_overrun.
3273 overwrite = local_read(&(cpu_buffer->overrun));
3276 * Here's the tricky part.
3278 * We need to move the pointer past the header page.
3279 * But we can only do that if a writer is not currently
3280 * moving it. The page before the header page has the
3281 * flag bit '1' set if it is pointing to the page we want.
3282 * but if the writer is in the process of moving it
3283 * than it will be '2' or already moved '0'.
3286 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3289 * If we did not convert it, then we must try again.
3295 * Yeah! We succeeded in replacing the page.
3297 * Now make the new head point back to the reader page.
3299 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3300 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3302 /* Finally update the reader page to the new head */
3303 cpu_buffer->reader_page = reader;
3304 rb_reset_reader_page(cpu_buffer);
3306 if (overwrite != cpu_buffer->last_overrun) {
3307 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3308 cpu_buffer->last_overrun = overwrite;
3314 arch_spin_unlock(&cpu_buffer->lock);
3315 local_irq_restore(flags);
3320 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3322 struct ring_buffer_event *event;
3323 struct buffer_page *reader;
3326 reader = rb_get_reader_page(cpu_buffer);
3328 /* This function should not be called when buffer is empty */
3329 if (RB_WARN_ON(cpu_buffer, !reader))
3332 event = rb_reader_event(cpu_buffer);
3334 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3337 rb_update_read_stamp(cpu_buffer, event);
3339 length = rb_event_length(event);
3340 cpu_buffer->reader_page->read += length;
3343 static void rb_advance_iter(struct ring_buffer_iter *iter)
3345 struct ring_buffer_per_cpu *cpu_buffer;
3346 struct ring_buffer_event *event;
3349 cpu_buffer = iter->cpu_buffer;
3352 * Check if we are at the end of the buffer.
3354 if (iter->head >= rb_page_size(iter->head_page)) {
3355 /* discarded commits can make the page empty */
3356 if (iter->head_page == cpu_buffer->commit_page)
3362 event = rb_iter_head_event(iter);
3364 length = rb_event_length(event);
3367 * This should not be called to advance the header if we are
3368 * at the tail of the buffer.
3370 if (RB_WARN_ON(cpu_buffer,
3371 (iter->head_page == cpu_buffer->commit_page) &&
3372 (iter->head + length > rb_commit_index(cpu_buffer))))
3375 rb_update_iter_read_stamp(iter, event);
3377 iter->head += length;
3379 /* check for end of page padding */
3380 if ((iter->head >= rb_page_size(iter->head_page)) &&
3381 (iter->head_page != cpu_buffer->commit_page))
3382 rb_advance_iter(iter);
3385 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3387 return cpu_buffer->lost_events;
3390 static struct ring_buffer_event *
3391 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3392 unsigned long *lost_events)
3394 struct ring_buffer_event *event;
3395 struct buffer_page *reader;
3400 * We repeat when a time extend is encountered.
3401 * Since the time extend is always attached to a data event,
3402 * we should never loop more than once.
3403 * (We never hit the following condition more than twice).
3405 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3408 reader = rb_get_reader_page(cpu_buffer);
3412 event = rb_reader_event(cpu_buffer);
3414 switch (event->type_len) {
3415 case RINGBUF_TYPE_PADDING:
3416 if (rb_null_event(event))
3417 RB_WARN_ON(cpu_buffer, 1);
3419 * Because the writer could be discarding every
3420 * event it creates (which would probably be bad)
3421 * if we were to go back to "again" then we may never
3422 * catch up, and will trigger the warn on, or lock
3423 * the box. Return the padding, and we will release
3424 * the current locks, and try again.
3428 case RINGBUF_TYPE_TIME_EXTEND:
3429 /* Internal data, OK to advance */
3430 rb_advance_reader(cpu_buffer);
3433 case RINGBUF_TYPE_TIME_STAMP:
3434 /* FIXME: not implemented */
3435 rb_advance_reader(cpu_buffer);
3438 case RINGBUF_TYPE_DATA:
3440 *ts = cpu_buffer->read_stamp + event->time_delta;
3441 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3442 cpu_buffer->cpu, ts);
3445 *lost_events = rb_lost_events(cpu_buffer);
3454 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3456 static struct ring_buffer_event *
3457 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3459 struct ring_buffer *buffer;
3460 struct ring_buffer_per_cpu *cpu_buffer;
3461 struct ring_buffer_event *event;
3464 cpu_buffer = iter->cpu_buffer;
3465 buffer = cpu_buffer->buffer;
3468 * Check if someone performed a consuming read to
3469 * the buffer. A consuming read invalidates the iterator
3470 * and we need to reset the iterator in this case.
3472 if (unlikely(iter->cache_read != cpu_buffer->read ||
3473 iter->cache_reader_page != cpu_buffer->reader_page))
3474 rb_iter_reset(iter);
3477 if (ring_buffer_iter_empty(iter))
3481 * We repeat when a time extend is encountered.
3482 * Since the time extend is always attached to a data event,
3483 * we should never loop more than once.
3484 * (We never hit the following condition more than twice).
3486 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3489 if (rb_per_cpu_empty(cpu_buffer))
3492 if (iter->head >= local_read(&iter->head_page->page->commit)) {
3497 event = rb_iter_head_event(iter);
3499 switch (event->type_len) {
3500 case RINGBUF_TYPE_PADDING:
3501 if (rb_null_event(event)) {
3505 rb_advance_iter(iter);
3508 case RINGBUF_TYPE_TIME_EXTEND:
3509 /* Internal data, OK to advance */
3510 rb_advance_iter(iter);
3513 case RINGBUF_TYPE_TIME_STAMP:
3514 /* FIXME: not implemented */
3515 rb_advance_iter(iter);
3518 case RINGBUF_TYPE_DATA:
3520 *ts = iter->read_stamp + event->time_delta;
3521 ring_buffer_normalize_time_stamp(buffer,
3522 cpu_buffer->cpu, ts);
3532 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3534 static inline int rb_ok_to_lock(void)
3537 * If an NMI die dumps out the content of the ring buffer
3538 * do not grab locks. We also permanently disable the ring
3539 * buffer too. A one time deal is all you get from reading
3540 * the ring buffer from an NMI.
3542 if (likely(!in_nmi()))
3545 tracing_off_permanent();
3550 * ring_buffer_peek - peek at the next event to be read
3551 * @buffer: The ring buffer to read
3552 * @cpu: The cpu to peak at
3553 * @ts: The timestamp counter of this event.
3554 * @lost_events: a variable to store if events were lost (may be NULL)
3556 * This will return the event that will be read next, but does
3557 * not consume the data.
3559 struct ring_buffer_event *
3560 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3561 unsigned long *lost_events)
3563 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3564 struct ring_buffer_event *event;
3565 unsigned long flags;
3568 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3571 dolock = rb_ok_to_lock();
3573 local_irq_save(flags);
3575 raw_spin_lock(&cpu_buffer->reader_lock);
3576 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3577 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3578 rb_advance_reader(cpu_buffer);
3580 raw_spin_unlock(&cpu_buffer->reader_lock);
3581 local_irq_restore(flags);
3583 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3590 * ring_buffer_iter_peek - peek at the next event to be read
3591 * @iter: The ring buffer iterator
3592 * @ts: The timestamp counter of this event.
3594 * This will return the event that will be read next, but does
3595 * not increment the iterator.
3597 struct ring_buffer_event *
3598 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3600 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3601 struct ring_buffer_event *event;
3602 unsigned long flags;
3605 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3606 event = rb_iter_peek(iter, ts);
3607 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3609 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3616 * ring_buffer_consume - return an event and consume it
3617 * @buffer: The ring buffer to get the next event from
3618 * @cpu: the cpu to read the buffer from
3619 * @ts: a variable to store the timestamp (may be NULL)
3620 * @lost_events: a variable to store if events were lost (may be NULL)
3622 * Returns the next event in the ring buffer, and that event is consumed.
3623 * Meaning, that sequential reads will keep returning a different event,
3624 * and eventually empty the ring buffer if the producer is slower.
3626 struct ring_buffer_event *
3627 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3628 unsigned long *lost_events)
3630 struct ring_buffer_per_cpu *cpu_buffer;
3631 struct ring_buffer_event *event = NULL;
3632 unsigned long flags;
3635 dolock = rb_ok_to_lock();
3638 /* might be called in atomic */
3641 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3644 cpu_buffer = buffer->buffers[cpu];
3645 local_irq_save(flags);
3647 raw_spin_lock(&cpu_buffer->reader_lock);
3649 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3651 cpu_buffer->lost_events = 0;
3652 rb_advance_reader(cpu_buffer);
3656 raw_spin_unlock(&cpu_buffer->reader_lock);
3657 local_irq_restore(flags);
3662 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3667 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3670 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3671 * @buffer: The ring buffer to read from
3672 * @cpu: The cpu buffer to iterate over
3674 * This performs the initial preparations necessary to iterate
3675 * through the buffer. Memory is allocated, buffer recording
3676 * is disabled, and the iterator pointer is returned to the caller.
3678 * Disabling buffer recordng prevents the reading from being
3679 * corrupted. This is not a consuming read, so a producer is not
3682 * After a sequence of ring_buffer_read_prepare calls, the user is
3683 * expected to make at least one call to ring_buffer_prepare_sync.
3684 * Afterwards, ring_buffer_read_start is invoked to get things going
3687 * This overall must be paired with ring_buffer_finish.
3689 struct ring_buffer_iter *
3690 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3692 struct ring_buffer_per_cpu *cpu_buffer;
3693 struct ring_buffer_iter *iter;
3695 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3698 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3702 cpu_buffer = buffer->buffers[cpu];
3704 iter->cpu_buffer = cpu_buffer;
3706 atomic_inc(&buffer->resize_disabled);
3707 atomic_inc(&cpu_buffer->record_disabled);
3711 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3714 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3716 * All previously invoked ring_buffer_read_prepare calls to prepare
3717 * iterators will be synchronized. Afterwards, read_buffer_read_start
3718 * calls on those iterators are allowed.
3721 ring_buffer_read_prepare_sync(void)
3723 synchronize_sched();
3725 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
3728 * ring_buffer_read_start - start a non consuming read of the buffer
3729 * @iter: The iterator returned by ring_buffer_read_prepare
3731 * This finalizes the startup of an iteration through the buffer.
3732 * The iterator comes from a call to ring_buffer_read_prepare and
3733 * an intervening ring_buffer_read_prepare_sync must have been
3736 * Must be paired with ring_buffer_finish.
3739 ring_buffer_read_start(struct ring_buffer_iter *iter)
3741 struct ring_buffer_per_cpu *cpu_buffer;
3742 unsigned long flags;
3747 cpu_buffer = iter->cpu_buffer;
3749 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3750 arch_spin_lock(&cpu_buffer->lock);
3751 rb_iter_reset(iter);
3752 arch_spin_unlock(&cpu_buffer->lock);
3753 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3755 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3758 * ring_buffer_finish - finish reading the iterator of the buffer
3759 * @iter: The iterator retrieved by ring_buffer_start
3761 * This re-enables the recording to the buffer, and frees the
3765 ring_buffer_read_finish(struct ring_buffer_iter *iter)
3767 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3770 * Ring buffer is disabled from recording, here's a good place
3771 * to check the integrity of the ring buffer.
3773 rb_check_pages(cpu_buffer);
3775 atomic_dec(&cpu_buffer->record_disabled);
3776 atomic_dec(&cpu_buffer->buffer->resize_disabled);
3779 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3782 * ring_buffer_read - read the next item in the ring buffer by the iterator
3783 * @iter: The ring buffer iterator
3784 * @ts: The time stamp of the event read.
3786 * This reads the next event in the ring buffer and increments the iterator.
3788 struct ring_buffer_event *
3789 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3791 struct ring_buffer_event *event;
3792 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3793 unsigned long flags;
3795 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3797 event = rb_iter_peek(iter, ts);
3801 if (event->type_len == RINGBUF_TYPE_PADDING)
3804 rb_advance_iter(iter);
3806 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3810 EXPORT_SYMBOL_GPL(ring_buffer_read);
3813 * ring_buffer_size - return the size of the ring buffer (in bytes)
3814 * @buffer: The ring buffer.
3816 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
3819 * Earlier, this method returned
3820 * BUF_PAGE_SIZE * buffer->nr_pages
3821 * Since the nr_pages field is now removed, we have converted this to
3822 * return the per cpu buffer value.
3824 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3827 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
3829 EXPORT_SYMBOL_GPL(ring_buffer_size);
3832 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3834 rb_head_page_deactivate(cpu_buffer);
3836 cpu_buffer->head_page
3837 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3838 local_set(&cpu_buffer->head_page->write, 0);
3839 local_set(&cpu_buffer->head_page->entries, 0);
3840 local_set(&cpu_buffer->head_page->page->commit, 0);
3842 cpu_buffer->head_page->read = 0;
3844 cpu_buffer->tail_page = cpu_buffer->head_page;
3845 cpu_buffer->commit_page = cpu_buffer->head_page;
3847 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3848 INIT_LIST_HEAD(&cpu_buffer->new_pages);
3849 local_set(&cpu_buffer->reader_page->write, 0);
3850 local_set(&cpu_buffer->reader_page->entries, 0);
3851 local_set(&cpu_buffer->reader_page->page->commit, 0);
3852 cpu_buffer->reader_page->read = 0;
3854 local_set(&cpu_buffer->commit_overrun, 0);
3855 local_set(&cpu_buffer->entries_bytes, 0);
3856 local_set(&cpu_buffer->overrun, 0);
3857 local_set(&cpu_buffer->entries, 0);
3858 local_set(&cpu_buffer->committing, 0);
3859 local_set(&cpu_buffer->commits, 0);
3860 cpu_buffer->read = 0;
3861 cpu_buffer->read_bytes = 0;
3863 cpu_buffer->write_stamp = 0;
3864 cpu_buffer->read_stamp = 0;
3866 cpu_buffer->lost_events = 0;
3867 cpu_buffer->last_overrun = 0;
3869 rb_head_page_activate(cpu_buffer);
3873 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3874 * @buffer: The ring buffer to reset a per cpu buffer of
3875 * @cpu: The CPU buffer to be reset
3877 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3879 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3880 unsigned long flags;
3882 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3885 atomic_inc(&buffer->resize_disabled);
3886 atomic_inc(&cpu_buffer->record_disabled);
3888 /* Make sure all commits have finished */
3889 synchronize_sched();
3891 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3893 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3896 arch_spin_lock(&cpu_buffer->lock);
3898 rb_reset_cpu(cpu_buffer);
3900 arch_spin_unlock(&cpu_buffer->lock);
3903 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3905 atomic_dec(&cpu_buffer->record_disabled);
3906 atomic_dec(&buffer->resize_disabled);
3908 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3911 * ring_buffer_reset - reset a ring buffer
3912 * @buffer: The ring buffer to reset all cpu buffers
3914 void ring_buffer_reset(struct ring_buffer *buffer)
3918 for_each_buffer_cpu(buffer, cpu)
3919 ring_buffer_reset_cpu(buffer, cpu);
3921 EXPORT_SYMBOL_GPL(ring_buffer_reset);
3924 * rind_buffer_empty - is the ring buffer empty?
3925 * @buffer: The ring buffer to test
3927 int ring_buffer_empty(struct ring_buffer *buffer)
3929 struct ring_buffer_per_cpu *cpu_buffer;
3930 unsigned long flags;
3935 dolock = rb_ok_to_lock();
3937 /* yes this is racy, but if you don't like the race, lock the buffer */
3938 for_each_buffer_cpu(buffer, cpu) {
3939 cpu_buffer = buffer->buffers[cpu];
3940 local_irq_save(flags);
3942 raw_spin_lock(&cpu_buffer->reader_lock);
3943 ret = rb_per_cpu_empty(cpu_buffer);
3945 raw_spin_unlock(&cpu_buffer->reader_lock);
3946 local_irq_restore(flags);
3954 EXPORT_SYMBOL_GPL(ring_buffer_empty);
3957 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3958 * @buffer: The ring buffer
3959 * @cpu: The CPU buffer to test
3961 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3963 struct ring_buffer_per_cpu *cpu_buffer;
3964 unsigned long flags;
3968 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3971 dolock = rb_ok_to_lock();
3973 cpu_buffer = buffer->buffers[cpu];
3974 local_irq_save(flags);
3976 raw_spin_lock(&cpu_buffer->reader_lock);
3977 ret = rb_per_cpu_empty(cpu_buffer);
3979 raw_spin_unlock(&cpu_buffer->reader_lock);
3980 local_irq_restore(flags);
3984 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
3986 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3988 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3989 * @buffer_a: One buffer to swap with
3990 * @buffer_b: The other buffer to swap with
3992 * This function is useful for tracers that want to take a "snapshot"
3993 * of a CPU buffer and has another back up buffer lying around.
3994 * it is expected that the tracer handles the cpu buffer not being
3995 * used at the moment.
3997 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
3998 struct ring_buffer *buffer_b, int cpu)
4000 struct ring_buffer_per_cpu *cpu_buffer_a;
4001 struct ring_buffer_per_cpu *cpu_buffer_b;
4004 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4005 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4008 cpu_buffer_a = buffer_a->buffers[cpu];
4009 cpu_buffer_b = buffer_b->buffers[cpu];
4011 /* At least make sure the two buffers are somewhat the same */
4012 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4017 if (ring_buffer_flags != RB_BUFFERS_ON)
4020 if (atomic_read(&buffer_a->record_disabled))
4023 if (atomic_read(&buffer_b->record_disabled))
4026 if (atomic_read(&cpu_buffer_a->record_disabled))
4029 if (atomic_read(&cpu_buffer_b->record_disabled))
4033 * We can't do a synchronize_sched here because this
4034 * function can be called in atomic context.
4035 * Normally this will be called from the same CPU as cpu.
4036 * If not it's up to the caller to protect this.
4038 atomic_inc(&cpu_buffer_a->record_disabled);
4039 atomic_inc(&cpu_buffer_b->record_disabled);
4042 if (local_read(&cpu_buffer_a->committing))
4044 if (local_read(&cpu_buffer_b->committing))
4047 buffer_a->buffers[cpu] = cpu_buffer_b;
4048 buffer_b->buffers[cpu] = cpu_buffer_a;
4050 cpu_buffer_b->buffer = buffer_a;
4051 cpu_buffer_a->buffer = buffer_b;
4056 atomic_dec(&cpu_buffer_a->record_disabled);
4057 atomic_dec(&cpu_buffer_b->record_disabled);
4061 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4062 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4065 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4066 * @buffer: the buffer to allocate for.
4068 * This function is used in conjunction with ring_buffer_read_page.
4069 * When reading a full page from the ring buffer, these functions
4070 * can be used to speed up the process. The calling function should
4071 * allocate a few pages first with this function. Then when it
4072 * needs to get pages from the ring buffer, it passes the result
4073 * of this function into ring_buffer_read_page, which will swap
4074 * the page that was allocated, with the read page of the buffer.
4077 * The page allocated, or NULL on error.
4079 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4081 struct buffer_data_page *bpage;
4084 page = alloc_pages_node(cpu_to_node(cpu),
4085 GFP_KERNEL | __GFP_NORETRY, 0);
4089 bpage = page_address(page);
4091 rb_init_page(bpage);
4095 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4098 * ring_buffer_free_read_page - free an allocated read page
4099 * @buffer: the buffer the page was allocate for
4100 * @data: the page to free
4102 * Free a page allocated from ring_buffer_alloc_read_page.
4104 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4106 free_page((unsigned long)data);
4108 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4111 * ring_buffer_read_page - extract a page from the ring buffer
4112 * @buffer: buffer to extract from
4113 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4114 * @len: amount to extract
4115 * @cpu: the cpu of the buffer to extract
4116 * @full: should the extraction only happen when the page is full.
4118 * This function will pull out a page from the ring buffer and consume it.
4119 * @data_page must be the address of the variable that was returned
4120 * from ring_buffer_alloc_read_page. This is because the page might be used
4121 * to swap with a page in the ring buffer.
4124 * rpage = ring_buffer_alloc_read_page(buffer);
4127 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4129 * process_page(rpage, ret);
4131 * When @full is set, the function will not return true unless
4132 * the writer is off the reader page.
4134 * Note: it is up to the calling functions to handle sleeps and wakeups.
4135 * The ring buffer can be used anywhere in the kernel and can not
4136 * blindly call wake_up. The layer that uses the ring buffer must be
4137 * responsible for that.
4140 * >=0 if data has been transferred, returns the offset of consumed data.
4141 * <0 if no data has been transferred.
4143 int ring_buffer_read_page(struct ring_buffer *buffer,
4144 void **data_page, size_t len, int cpu, int full)
4146 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4147 struct ring_buffer_event *event;
4148 struct buffer_data_page *bpage;
4149 struct buffer_page *reader;
4150 unsigned long missed_events;
4151 unsigned long flags;
4152 unsigned int commit;
4157 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4161 * If len is not big enough to hold the page header, then
4162 * we can not copy anything.
4164 if (len <= BUF_PAGE_HDR_SIZE)
4167 len -= BUF_PAGE_HDR_SIZE;
4176 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4178 reader = rb_get_reader_page(cpu_buffer);
4182 event = rb_reader_event(cpu_buffer);
4184 read = reader->read;
4185 commit = rb_page_commit(reader);
4187 /* Check if any events were dropped */
4188 missed_events = cpu_buffer->lost_events;
4191 * If this page has been partially read or
4192 * if len is not big enough to read the rest of the page or
4193 * a writer is still on the page, then
4194 * we must copy the data from the page to the buffer.
4195 * Otherwise, we can simply swap the page with the one passed in.
4197 if (read || (len < (commit - read)) ||
4198 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4199 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4200 unsigned int rpos = read;
4201 unsigned int pos = 0;
4207 if (len > (commit - read))
4208 len = (commit - read);
4210 /* Always keep the time extend and data together */
4211 size = rb_event_ts_length(event);
4216 /* save the current timestamp, since the user will need it */
4217 save_timestamp = cpu_buffer->read_stamp;
4219 /* Need to copy one event at a time */
4221 /* We need the size of one event, because
4222 * rb_advance_reader only advances by one event,
4223 * whereas rb_event_ts_length may include the size of
4224 * one or two events.
4225 * We have already ensured there's enough space if this
4226 * is a time extend. */
4227 size = rb_event_length(event);
4228 memcpy(bpage->data + pos, rpage->data + rpos, size);
4232 rb_advance_reader(cpu_buffer);
4233 rpos = reader->read;
4239 event = rb_reader_event(cpu_buffer);
4240 /* Always keep the time extend and data together */
4241 size = rb_event_ts_length(event);
4242 } while (len >= size);
4245 local_set(&bpage->commit, pos);
4246 bpage->time_stamp = save_timestamp;
4248 /* we copied everything to the beginning */
4251 /* update the entry counter */
4252 cpu_buffer->read += rb_page_entries(reader);
4253 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4255 /* swap the pages */
4256 rb_init_page(bpage);
4257 bpage = reader->page;
4258 reader->page = *data_page;
4259 local_set(&reader->write, 0);
4260 local_set(&reader->entries, 0);
4265 * Use the real_end for the data size,
4266 * This gives us a chance to store the lost events
4269 if (reader->real_end)
4270 local_set(&bpage->commit, reader->real_end);
4274 cpu_buffer->lost_events = 0;
4276 commit = local_read(&bpage->commit);
4278 * Set a flag in the commit field if we lost events
4280 if (missed_events) {
4281 /* If there is room at the end of the page to save the
4282 * missed events, then record it there.
4284 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4285 memcpy(&bpage->data[commit], &missed_events,
4286 sizeof(missed_events));
4287 local_add(RB_MISSED_STORED, &bpage->commit);
4288 commit += sizeof(missed_events);
4290 local_add(RB_MISSED_EVENTS, &bpage->commit);
4294 * This page may be off to user land. Zero it out here.
4296 if (commit < BUF_PAGE_SIZE)
4297 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4300 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4305 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4307 #ifdef CONFIG_HOTPLUG_CPU
4308 static int rb_cpu_notify(struct notifier_block *self,
4309 unsigned long action, void *hcpu)
4311 struct ring_buffer *buffer =
4312 container_of(self, struct ring_buffer, cpu_notify);
4313 long cpu = (long)hcpu;
4314 int cpu_i, nr_pages_same;
4315 unsigned int nr_pages;
4318 case CPU_UP_PREPARE:
4319 case CPU_UP_PREPARE_FROZEN:
4320 if (cpumask_test_cpu(cpu, buffer->cpumask))
4325 /* check if all cpu sizes are same */
4326 for_each_buffer_cpu(buffer, cpu_i) {
4327 /* fill in the size from first enabled cpu */
4329 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4330 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4335 /* allocate minimum pages, user can later expand it */
4338 buffer->buffers[cpu] =
4339 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4340 if (!buffer->buffers[cpu]) {
4341 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4346 cpumask_set_cpu(cpu, buffer->cpumask);
4348 case CPU_DOWN_PREPARE:
4349 case CPU_DOWN_PREPARE_FROZEN:
4352 * If we were to free the buffer, then the user would
4353 * lose any trace that was in the buffer.