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);
1078 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1080 ret = rb_allocate_pages(cpu_buffer, nr_pages);
1082 goto fail_free_reader;
1084 cpu_buffer->head_page
1085 = list_entry(cpu_buffer->pages, struct buffer_page, list);
1086 cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
1088 rb_head_page_activate(cpu_buffer);
1093 free_buffer_page(cpu_buffer->reader_page);
1100 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
1102 struct list_head *head = cpu_buffer->pages;
1103 struct buffer_page *bpage, *tmp;
1105 free_buffer_page(cpu_buffer->reader_page);
1107 rb_head_page_deactivate(cpu_buffer);
1110 list_for_each_entry_safe(bpage, tmp, head, list) {
1111 list_del_init(&bpage->list);
1112 free_buffer_page(bpage);
1114 bpage = list_entry(head, struct buffer_page, list);
1115 free_buffer_page(bpage);
1121 #ifdef CONFIG_HOTPLUG_CPU
1122 static int rb_cpu_notify(struct notifier_block *self,
1123 unsigned long action, void *hcpu);
1127 * ring_buffer_alloc - allocate a new ring_buffer
1128 * @size: the size in bytes per cpu that is needed.
1129 * @flags: attributes to set for the ring buffer.
1131 * Currently the only flag that is available is the RB_FL_OVERWRITE
1132 * flag. This flag means that the buffer will overwrite old data
1133 * when the buffer wraps. If this flag is not set, the buffer will
1134 * drop data when the tail hits the head.
1136 struct ring_buffer *__ring_buffer_alloc(unsigned long size, unsigned flags,
1137 struct lock_class_key *key)
1139 struct ring_buffer *buffer;
1143 /* keep it in its own cache line */
1144 buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
1149 if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
1150 goto fail_free_buffer;
1152 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1153 buffer->flags = flags;
1154 buffer->clock = trace_clock_local;
1155 buffer->reader_lock_key = key;
1157 /* need at least two pages */
1162 * In case of non-hotplug cpu, if the ring-buffer is allocated
1163 * in early initcall, it will not be notified of secondary cpus.
1164 * In that off case, we need to allocate for all possible cpus.
1166 #ifdef CONFIG_HOTPLUG_CPU
1168 cpumask_copy(buffer->cpumask, cpu_online_mask);
1170 cpumask_copy(buffer->cpumask, cpu_possible_mask);
1172 buffer->cpus = nr_cpu_ids;
1174 bsize = sizeof(void *) * nr_cpu_ids;
1175 buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
1177 if (!buffer->buffers)
1178 goto fail_free_cpumask;
1180 for_each_buffer_cpu(buffer, cpu) {
1181 buffer->buffers[cpu] =
1182 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
1183 if (!buffer->buffers[cpu])
1184 goto fail_free_buffers;
1187 #ifdef CONFIG_HOTPLUG_CPU
1188 buffer->cpu_notify.notifier_call = rb_cpu_notify;
1189 buffer->cpu_notify.priority = 0;
1190 register_cpu_notifier(&buffer->cpu_notify);
1194 mutex_init(&buffer->mutex);
1199 for_each_buffer_cpu(buffer, cpu) {
1200 if (buffer->buffers[cpu])
1201 rb_free_cpu_buffer(buffer->buffers[cpu]);
1203 kfree(buffer->buffers);
1206 free_cpumask_var(buffer->cpumask);
1213 EXPORT_SYMBOL_GPL(__ring_buffer_alloc);
1216 * ring_buffer_free - free a ring buffer.
1217 * @buffer: the buffer to free.
1220 ring_buffer_free(struct ring_buffer *buffer)
1226 #ifdef CONFIG_HOTPLUG_CPU
1227 unregister_cpu_notifier(&buffer->cpu_notify);
1230 for_each_buffer_cpu(buffer, cpu)
1231 rb_free_cpu_buffer(buffer->buffers[cpu]);
1235 kfree(buffer->buffers);
1236 free_cpumask_var(buffer->cpumask);
1240 EXPORT_SYMBOL_GPL(ring_buffer_free);
1242 void ring_buffer_set_clock(struct ring_buffer *buffer,
1245 buffer->clock = clock;
1248 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
1250 static inline unsigned long rb_page_entries(struct buffer_page *bpage)
1252 return local_read(&bpage->entries) & RB_WRITE_MASK;
1255 static inline unsigned long rb_page_write(struct buffer_page *bpage)
1257 return local_read(&bpage->write) & RB_WRITE_MASK;
1261 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned int nr_pages)
1263 struct list_head *tail_page, *to_remove, *next_page;
1264 struct buffer_page *to_remove_page, *tmp_iter_page;
1265 struct buffer_page *last_page, *first_page;
1266 unsigned int nr_removed;
1267 unsigned long head_bit;
1272 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1273 atomic_inc(&cpu_buffer->record_disabled);
1275 * We don't race with the readers since we have acquired the reader
1276 * lock. We also don't race with writers after disabling recording.
1277 * This makes it easy to figure out the first and the last page to be
1278 * removed from the list. We unlink all the pages in between including
1279 * the first and last pages. This is done in a busy loop so that we
1280 * lose the least number of traces.
1281 * The pages are freed after we restart recording and unlock readers.
1283 tail_page = &cpu_buffer->tail_page->list;
1286 * tail page might be on reader page, we remove the next page
1287 * from the ring buffer
1289 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
1290 tail_page = rb_list_head(tail_page->next);
1291 to_remove = tail_page;
1293 /* start of pages to remove */
1294 first_page = list_entry(rb_list_head(to_remove->next),
1295 struct buffer_page, list);
1297 for (nr_removed = 0; nr_removed < nr_pages; nr_removed++) {
1298 to_remove = rb_list_head(to_remove)->next;
1299 head_bit |= (unsigned long)to_remove & RB_PAGE_HEAD;
1302 next_page = rb_list_head(to_remove)->next;
1305 * Now we remove all pages between tail_page and next_page.
1306 * Make sure that we have head_bit value preserved for the
1309 tail_page->next = (struct list_head *)((unsigned long)next_page |
1311 next_page = rb_list_head(next_page);
1312 next_page->prev = tail_page;
1314 /* make sure pages points to a valid page in the ring buffer */
1315 cpu_buffer->pages = next_page;
1317 /* update head page */
1319 cpu_buffer->head_page = list_entry(next_page,
1320 struct buffer_page, list);
1323 * change read pointer to make sure any read iterators reset
1326 cpu_buffer->read = 0;
1328 /* pages are removed, resume tracing and then free the pages */
1329 atomic_dec(&cpu_buffer->record_disabled);
1330 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1332 RB_WARN_ON(cpu_buffer, list_empty(cpu_buffer->pages));
1334 /* last buffer page to remove */
1335 last_page = list_entry(rb_list_head(to_remove), struct buffer_page,
1337 tmp_iter_page = first_page;
1340 to_remove_page = tmp_iter_page;
1341 rb_inc_page(cpu_buffer, &tmp_iter_page);
1343 /* update the counters */
1344 page_entries = rb_page_entries(to_remove_page);
1347 * If something was added to this page, it was full
1348 * since it is not the tail page. So we deduct the
1349 * bytes consumed in ring buffer from here.
1350 * No need to update overruns, since this page is
1351 * deleted from ring buffer and its entries are
1352 * already accounted for.
1354 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1358 * We have already removed references to this list item, just
1359 * free up the buffer_page and its page
1361 free_buffer_page(to_remove_page);
1364 } while (to_remove_page != last_page);
1366 RB_WARN_ON(cpu_buffer, nr_removed);
1368 return nr_removed == 0;
1372 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1374 struct list_head *pages = &cpu_buffer->new_pages;
1375 int retries, success;
1377 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1379 * We are holding the reader lock, so the reader page won't be swapped
1380 * in the ring buffer. Now we are racing with the writer trying to
1381 * move head page and the tail page.
1382 * We are going to adapt the reader page update process where:
1383 * 1. We first splice the start and end of list of new pages between
1384 * the head page and its previous page.
1385 * 2. We cmpxchg the prev_page->next to point from head page to the
1386 * start of new pages list.
1387 * 3. Finally, we update the head->prev to the end of new list.
1389 * We will try this process 10 times, to make sure that we don't keep
1395 struct list_head *head_page, *prev_page, *r;
1396 struct list_head *last_page, *first_page;
1397 struct list_head *head_page_with_bit;
1399 head_page = &rb_set_head_page(cpu_buffer)->list;
1400 prev_page = head_page->prev;
1402 first_page = pages->next;
1403 last_page = pages->prev;
1405 head_page_with_bit = (struct list_head *)
1406 ((unsigned long)head_page | RB_PAGE_HEAD);
1408 last_page->next = head_page_with_bit;
1409 first_page->prev = prev_page;
1411 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1413 if (r == head_page_with_bit) {
1415 * yay, we replaced the page pointer to our new list,
1416 * now, we just have to update to head page's prev
1417 * pointer to point to end of list
1419 head_page->prev = last_page;
1426 INIT_LIST_HEAD(pages);
1428 * If we weren't successful in adding in new pages, warn and stop
1431 RB_WARN_ON(cpu_buffer, !success);
1432 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1434 /* free pages if they weren't inserted */
1436 struct buffer_page *bpage, *tmp;
1437 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1439 list_del_init(&bpage->list);
1440 free_buffer_page(bpage);
1446 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1450 if (cpu_buffer->nr_pages_to_update > 0)
1451 success = rb_insert_pages(cpu_buffer);
1453 success = rb_remove_pages(cpu_buffer,
1454 -cpu_buffer->nr_pages_to_update);
1457 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1460 static void update_pages_handler(struct work_struct *work)
1462 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1463 struct ring_buffer_per_cpu, update_pages_work);
1464 rb_update_pages(cpu_buffer);
1465 complete(&cpu_buffer->update_done);
1469 * ring_buffer_resize - resize the ring buffer
1470 * @buffer: the buffer to resize.
1471 * @size: the new size.
1473 * Minimum size is 2 * BUF_PAGE_SIZE.
1475 * Returns 0 on success and < 0 on failure.
1477 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1480 struct ring_buffer_per_cpu *cpu_buffer;
1485 * Always succeed at resizing a non-existent buffer:
1490 /* Make sure the requested buffer exists */
1491 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1492 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1495 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1496 size *= BUF_PAGE_SIZE;
1498 /* we need a minimum of two pages */
1499 if (size < BUF_PAGE_SIZE * 2)
1500 size = BUF_PAGE_SIZE * 2;
1502 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1505 * Don't succeed if resizing is disabled, as a reader might be
1506 * manipulating the ring buffer and is expecting a sane state while
1509 if (atomic_read(&buffer->resize_disabled))
1512 /* prevent another thread from changing buffer sizes */
1513 mutex_lock(&buffer->mutex);
1515 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1516 /* calculate the pages to update */
1517 for_each_buffer_cpu(buffer, cpu) {
1518 cpu_buffer = buffer->buffers[cpu];
1520 cpu_buffer->nr_pages_to_update = nr_pages -
1521 cpu_buffer->nr_pages;
1523 * nothing more to do for removing pages or no update
1525 if (cpu_buffer->nr_pages_to_update <= 0)
1528 * to add pages, make sure all new pages can be
1529 * allocated without receiving ENOMEM
1531 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1532 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1533 &cpu_buffer->new_pages, cpu)) {
1534 /* not enough memory for new pages */
1542 * Fire off all the required work handlers
1543 * We can't schedule on offline CPUs, but it's not necessary
1544 * since we can change their buffer sizes without any race.
1546 for_each_buffer_cpu(buffer, cpu) {
1547 cpu_buffer = buffer->buffers[cpu];
1548 if (!cpu_buffer->nr_pages_to_update)
1551 if (cpu_online(cpu))
1552 schedule_work_on(cpu,
1553 &cpu_buffer->update_pages_work);
1555 rb_update_pages(cpu_buffer);
1558 /* wait for all the updates to complete */
1559 for_each_buffer_cpu(buffer, cpu) {
1560 cpu_buffer = buffer->buffers[cpu];
1561 if (!cpu_buffer->nr_pages_to_update)
1564 if (cpu_online(cpu))
1565 wait_for_completion(&cpu_buffer->update_done);
1566 cpu_buffer->nr_pages_to_update = 0;
1571 cpu_buffer = buffer->buffers[cpu_id];
1573 if (nr_pages == cpu_buffer->nr_pages)
1576 cpu_buffer->nr_pages_to_update = nr_pages -
1577 cpu_buffer->nr_pages;
1579 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1580 if (cpu_buffer->nr_pages_to_update > 0 &&
1581 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1582 &cpu_buffer->new_pages, cpu_id)) {
1589 if (cpu_online(cpu_id)) {
1590 schedule_work_on(cpu_id,
1591 &cpu_buffer->update_pages_work);
1592 wait_for_completion(&cpu_buffer->update_done);
1594 rb_update_pages(cpu_buffer);
1596 cpu_buffer->nr_pages_to_update = 0;
1602 * The ring buffer resize can happen with the ring buffer
1603 * enabled, so that the update disturbs the tracing as little
1604 * as possible. But if the buffer is disabled, we do not need
1605 * to worry about that, and we can take the time to verify
1606 * that the buffer is not corrupt.
1608 if (atomic_read(&buffer->record_disabled)) {
1609 atomic_inc(&buffer->record_disabled);
1611 * Even though the buffer was disabled, we must make sure
1612 * that it is truly disabled before calling rb_check_pages.
1613 * There could have been a race between checking
1614 * record_disable and incrementing it.
1616 synchronize_sched();
1617 for_each_buffer_cpu(buffer, cpu) {
1618 cpu_buffer = buffer->buffers[cpu];
1619 rb_check_pages(cpu_buffer);
1621 atomic_dec(&buffer->record_disabled);
1624 mutex_unlock(&buffer->mutex);
1628 for_each_buffer_cpu(buffer, cpu) {
1629 struct buffer_page *bpage, *tmp;
1631 cpu_buffer = buffer->buffers[cpu];
1632 cpu_buffer->nr_pages_to_update = 0;
1634 if (list_empty(&cpu_buffer->new_pages))
1637 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1639 list_del_init(&bpage->list);
1640 free_buffer_page(bpage);
1643 mutex_unlock(&buffer->mutex);
1646 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1648 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1650 mutex_lock(&buffer->mutex);
1652 buffer->flags |= RB_FL_OVERWRITE;
1654 buffer->flags &= ~RB_FL_OVERWRITE;
1655 mutex_unlock(&buffer->mutex);
1657 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1659 static inline void *
1660 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1662 return bpage->data + index;
1665 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1667 return bpage->page->data + index;
1670 static inline struct ring_buffer_event *
1671 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1673 return __rb_page_index(cpu_buffer->reader_page,
1674 cpu_buffer->reader_page->read);
1677 static inline struct ring_buffer_event *
1678 rb_iter_head_event(struct ring_buffer_iter *iter)
1680 return __rb_page_index(iter->head_page, iter->head);
1683 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1685 return local_read(&bpage->page->commit);
1688 /* Size is determined by what has been committed */
1689 static inline unsigned rb_page_size(struct buffer_page *bpage)
1691 return rb_page_commit(bpage);
1694 static inline unsigned
1695 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1697 return rb_page_commit(cpu_buffer->commit_page);
1700 static inline unsigned
1701 rb_event_index(struct ring_buffer_event *event)
1703 unsigned long addr = (unsigned long)event;
1705 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1709 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1710 struct ring_buffer_event *event)
1712 unsigned long addr = (unsigned long)event;
1713 unsigned long index;
1715 index = rb_event_index(event);
1718 return cpu_buffer->commit_page->page == (void *)addr &&
1719 rb_commit_index(cpu_buffer) == index;
1723 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1725 unsigned long max_count;
1728 * We only race with interrupts and NMIs on this CPU.
1729 * If we own the commit event, then we can commit
1730 * all others that interrupted us, since the interruptions
1731 * are in stack format (they finish before they come
1732 * back to us). This allows us to do a simple loop to
1733 * assign the commit to the tail.
1736 max_count = cpu_buffer->nr_pages * 100;
1738 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1739 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1741 if (RB_WARN_ON(cpu_buffer,
1742 rb_is_reader_page(cpu_buffer->tail_page)))
1744 local_set(&cpu_buffer->commit_page->page->commit,
1745 rb_page_write(cpu_buffer->commit_page));
1746 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1747 cpu_buffer->write_stamp =
1748 cpu_buffer->commit_page->page->time_stamp;
1749 /* add barrier to keep gcc from optimizing too much */
1752 while (rb_commit_index(cpu_buffer) !=
1753 rb_page_write(cpu_buffer->commit_page)) {
1755 local_set(&cpu_buffer->commit_page->page->commit,
1756 rb_page_write(cpu_buffer->commit_page));
1757 RB_WARN_ON(cpu_buffer,
1758 local_read(&cpu_buffer->commit_page->page->commit) &
1763 /* again, keep gcc from optimizing */
1767 * If an interrupt came in just after the first while loop
1768 * and pushed the tail page forward, we will be left with
1769 * a dangling commit that will never go forward.
1771 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1775 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1777 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1778 cpu_buffer->reader_page->read = 0;
1781 static void rb_inc_iter(struct ring_buffer_iter *iter)
1783 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1786 * The iterator could be on the reader page (it starts there).
1787 * But the head could have moved, since the reader was
1788 * found. Check for this case and assign the iterator
1789 * to the head page instead of next.
1791 if (iter->head_page == cpu_buffer->reader_page)
1792 iter->head_page = rb_set_head_page(cpu_buffer);
1794 rb_inc_page(cpu_buffer, &iter->head_page);
1796 iter->read_stamp = iter->head_page->page->time_stamp;
1800 /* Slow path, do not inline */
1801 static noinline struct ring_buffer_event *
1802 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1804 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1806 /* Not the first event on the page? */
1807 if (rb_event_index(event)) {
1808 event->time_delta = delta & TS_MASK;
1809 event->array[0] = delta >> TS_SHIFT;
1811 /* nope, just zero it */
1812 event->time_delta = 0;
1813 event->array[0] = 0;
1816 return skip_time_extend(event);
1820 * ring_buffer_update_event - update event type and data
1821 * @event: the even to update
1822 * @type: the type of event
1823 * @length: the size of the event field in the ring buffer
1825 * Update the type and data fields of the event. The length
1826 * is the actual size that is written to the ring buffer,
1827 * and with this, we can determine what to place into the
1831 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1832 struct ring_buffer_event *event, unsigned length,
1833 int add_timestamp, u64 delta)
1835 /* Only a commit updates the timestamp */
1836 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
1840 * If we need to add a timestamp, then we
1841 * add it to the start of the resevered space.
1843 if (unlikely(add_timestamp)) {
1844 event = rb_add_time_stamp(event, delta);
1845 length -= RB_LEN_TIME_EXTEND;
1849 event->time_delta = delta;
1850 length -= RB_EVNT_HDR_SIZE;
1851 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
1852 event->type_len = 0;
1853 event->array[0] = length;
1855 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1859 * rb_handle_head_page - writer hit the head page
1861 * Returns: +1 to retry page
1866 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1867 struct buffer_page *tail_page,
1868 struct buffer_page *next_page)
1870 struct buffer_page *new_head;
1875 entries = rb_page_entries(next_page);
1878 * The hard part is here. We need to move the head
1879 * forward, and protect against both readers on
1880 * other CPUs and writers coming in via interrupts.
1882 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1886 * type can be one of four:
1887 * NORMAL - an interrupt already moved it for us
1888 * HEAD - we are the first to get here.
1889 * UPDATE - we are the interrupt interrupting
1891 * MOVED - a reader on another CPU moved the next
1892 * pointer to its reader page. Give up
1899 * We changed the head to UPDATE, thus
1900 * it is our responsibility to update
1903 local_add(entries, &cpu_buffer->overrun);
1904 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1907 * The entries will be zeroed out when we move the
1911 /* still more to do */
1914 case RB_PAGE_UPDATE:
1916 * This is an interrupt that interrupt the
1917 * previous update. Still more to do.
1920 case RB_PAGE_NORMAL:
1922 * An interrupt came in before the update
1923 * and processed this for us.
1924 * Nothing left to do.
1929 * The reader is on another CPU and just did
1930 * a swap with our next_page.
1935 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1940 * Now that we are here, the old head pointer is
1941 * set to UPDATE. This will keep the reader from
1942 * swapping the head page with the reader page.
1943 * The reader (on another CPU) will spin till
1946 * We just need to protect against interrupts
1947 * doing the job. We will set the next pointer
1948 * to HEAD. After that, we set the old pointer
1949 * to NORMAL, but only if it was HEAD before.
1950 * otherwise we are an interrupt, and only
1951 * want the outer most commit to reset it.
1953 new_head = next_page;
1954 rb_inc_page(cpu_buffer, &new_head);
1956 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1960 * Valid returns are:
1961 * HEAD - an interrupt came in and already set it.
1962 * NORMAL - One of two things:
1963 * 1) We really set it.
1964 * 2) A bunch of interrupts came in and moved
1965 * the page forward again.
1969 case RB_PAGE_NORMAL:
1973 RB_WARN_ON(cpu_buffer, 1);
1978 * It is possible that an interrupt came in,
1979 * set the head up, then more interrupts came in
1980 * and moved it again. When we get back here,
1981 * the page would have been set to NORMAL but we
1982 * just set it back to HEAD.
1984 * How do you detect this? Well, if that happened
1985 * the tail page would have moved.
1987 if (ret == RB_PAGE_NORMAL) {
1989 * If the tail had moved passed next, then we need
1990 * to reset the pointer.
1992 if (cpu_buffer->tail_page != tail_page &&
1993 cpu_buffer->tail_page != next_page)
1994 rb_head_page_set_normal(cpu_buffer, new_head,
2000 * If this was the outer most commit (the one that
2001 * changed the original pointer from HEAD to UPDATE),
2002 * then it is up to us to reset it to NORMAL.
2004 if (type == RB_PAGE_HEAD) {
2005 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2008 if (RB_WARN_ON(cpu_buffer,
2009 ret != RB_PAGE_UPDATE))
2016 static unsigned rb_calculate_event_length(unsigned length)
2018 struct ring_buffer_event event; /* Used only for sizeof array */
2020 /* zero length can cause confusions */
2024 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2025 length += sizeof(event.array[0]);
2027 length += RB_EVNT_HDR_SIZE;
2028 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2034 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2035 struct buffer_page *tail_page,
2036 unsigned long tail, unsigned long length)
2038 struct ring_buffer_event *event;
2041 * Only the event that crossed the page boundary
2042 * must fill the old tail_page with padding.
2044 if (tail >= BUF_PAGE_SIZE) {
2046 * If the page was filled, then we still need
2047 * to update the real_end. Reset it to zero
2048 * and the reader will ignore it.
2050 if (tail == BUF_PAGE_SIZE)
2051 tail_page->real_end = 0;
2053 local_sub(length, &tail_page->write);
2057 event = __rb_page_index(tail_page, tail);
2058 kmemcheck_annotate_bitfield(event, bitfield);
2060 /* account for padding bytes */
2061 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2064 * Save the original length to the meta data.
2065 * This will be used by the reader to add lost event
2068 tail_page->real_end = tail;
2071 * If this event is bigger than the minimum size, then
2072 * we need to be careful that we don't subtract the
2073 * write counter enough to allow another writer to slip
2075 * We put in a discarded commit instead, to make sure
2076 * that this space is not used again.
2078 * If we are less than the minimum size, we don't need to
2081 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2082 /* No room for any events */
2084 /* Mark the rest of the page with padding */
2085 rb_event_set_padding(event);
2087 /* Set the write back to the previous setting */
2088 local_sub(length, &tail_page->write);
2092 /* Put in a discarded event */
2093 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2094 event->type_len = RINGBUF_TYPE_PADDING;
2095 /* time delta must be non zero */
2096 event->time_delta = 1;
2098 /* Set write to end of buffer */
2099 length = (tail + length) - BUF_PAGE_SIZE;
2100 local_sub(length, &tail_page->write);
2104 * This is the slow path, force gcc not to inline it.
2106 static noinline struct ring_buffer_event *
2107 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2108 unsigned long length, unsigned long tail,
2109 struct buffer_page *tail_page, u64 ts)
2111 struct buffer_page *commit_page = cpu_buffer->commit_page;
2112 struct ring_buffer *buffer = cpu_buffer->buffer;
2113 struct buffer_page *next_page;
2116 next_page = tail_page;
2118 rb_inc_page(cpu_buffer, &next_page);
2121 * If for some reason, we had an interrupt storm that made
2122 * it all the way around the buffer, bail, and warn
2125 if (unlikely(next_page == commit_page)) {
2126 local_inc(&cpu_buffer->commit_overrun);
2131 * This is where the fun begins!
2133 * We are fighting against races between a reader that
2134 * could be on another CPU trying to swap its reader
2135 * page with the buffer head.
2137 * We are also fighting against interrupts coming in and
2138 * moving the head or tail on us as well.
2140 * If the next page is the head page then we have filled
2141 * the buffer, unless the commit page is still on the
2144 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2147 * If the commit is not on the reader page, then
2148 * move the header page.
2150 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2152 * If we are not in overwrite mode,
2153 * this is easy, just stop here.
2155 if (!(buffer->flags & RB_FL_OVERWRITE))
2158 ret = rb_handle_head_page(cpu_buffer,
2167 * We need to be careful here too. The
2168 * commit page could still be on the reader
2169 * page. We could have a small buffer, and
2170 * have filled up the buffer with events
2171 * from interrupts and such, and wrapped.
2173 * Note, if the tail page is also the on the
2174 * reader_page, we let it move out.
2176 if (unlikely((cpu_buffer->commit_page !=
2177 cpu_buffer->tail_page) &&
2178 (cpu_buffer->commit_page ==
2179 cpu_buffer->reader_page))) {
2180 local_inc(&cpu_buffer->commit_overrun);
2186 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2189 * Nested commits always have zero deltas, so
2190 * just reread the time stamp
2192 ts = rb_time_stamp(buffer);
2193 next_page->page->time_stamp = ts;
2198 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2200 /* fail and let the caller try again */
2201 return ERR_PTR(-EAGAIN);
2205 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2210 static struct ring_buffer_event *
2211 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2212 unsigned long length, u64 ts,
2213 u64 delta, int add_timestamp)
2215 struct buffer_page *tail_page;
2216 struct ring_buffer_event *event;
2217 unsigned long tail, write;
2220 * If the time delta since the last event is too big to
2221 * hold in the time field of the event, then we append a
2222 * TIME EXTEND event ahead of the data event.
2224 if (unlikely(add_timestamp))
2225 length += RB_LEN_TIME_EXTEND;
2227 tail_page = cpu_buffer->tail_page;
2228 write = local_add_return(length, &tail_page->write);
2230 /* set write to only the index of the write */
2231 write &= RB_WRITE_MASK;
2232 tail = write - length;
2234 /* See if we shot pass the end of this buffer page */
2235 if (unlikely(write > BUF_PAGE_SIZE))
2236 return rb_move_tail(cpu_buffer, length, tail,
2239 /* We reserved something on the buffer */
2241 event = __rb_page_index(tail_page, tail);
2242 kmemcheck_annotate_bitfield(event, bitfield);
2243 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2245 local_inc(&tail_page->entries);
2248 * If this is the first commit on the page, then update
2252 tail_page->page->time_stamp = ts;
2254 /* account for these added bytes */
2255 local_add(length, &cpu_buffer->entries_bytes);
2261 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2262 struct ring_buffer_event *event)
2264 unsigned long new_index, old_index;
2265 struct buffer_page *bpage;
2266 unsigned long index;
2269 new_index = rb_event_index(event);
2270 old_index = new_index + rb_event_ts_length(event);
2271 addr = (unsigned long)event;
2274 bpage = cpu_buffer->tail_page;
2276 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2277 unsigned long write_mask =
2278 local_read(&bpage->write) & ~RB_WRITE_MASK;
2279 unsigned long event_length = rb_event_length(event);
2281 * This is on the tail page. It is possible that
2282 * a write could come in and move the tail page
2283 * and write to the next page. That is fine
2284 * because we just shorten what is on this page.
2286 old_index += write_mask;
2287 new_index += write_mask;
2288 index = local_cmpxchg(&bpage->write, old_index, new_index);
2289 if (index == old_index) {
2290 /* update counters */
2291 local_sub(event_length, &cpu_buffer->entries_bytes);
2296 /* could not discard */
2300 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2302 local_inc(&cpu_buffer->committing);
2303 local_inc(&cpu_buffer->commits);
2306 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2308 unsigned long commits;
2310 if (RB_WARN_ON(cpu_buffer,
2311 !local_read(&cpu_buffer->committing)))
2315 commits = local_read(&cpu_buffer->commits);
2316 /* synchronize with interrupts */
2318 if (local_read(&cpu_buffer->committing) == 1)
2319 rb_set_commit_to_write(cpu_buffer);
2321 local_dec(&cpu_buffer->committing);
2323 /* synchronize with interrupts */
2327 * Need to account for interrupts coming in between the
2328 * updating of the commit page and the clearing of the
2329 * committing counter.
2331 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2332 !local_read(&cpu_buffer->committing)) {
2333 local_inc(&cpu_buffer->committing);
2338 static struct ring_buffer_event *
2339 rb_reserve_next_event(struct ring_buffer *buffer,
2340 struct ring_buffer_per_cpu *cpu_buffer,
2341 unsigned long length)
2343 struct ring_buffer_event *event;
2349 rb_start_commit(cpu_buffer);
2351 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2353 * Due to the ability to swap a cpu buffer from a buffer
2354 * it is possible it was swapped before we committed.
2355 * (committing stops a swap). We check for it here and
2356 * if it happened, we have to fail the write.
2359 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2360 local_dec(&cpu_buffer->committing);
2361 local_dec(&cpu_buffer->commits);
2366 length = rb_calculate_event_length(length);
2372 * We allow for interrupts to reenter here and do a trace.
2373 * If one does, it will cause this original code to loop
2374 * back here. Even with heavy interrupts happening, this
2375 * should only happen a few times in a row. If this happens
2376 * 1000 times in a row, there must be either an interrupt
2377 * storm or we have something buggy.
2380 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2383 ts = rb_time_stamp(cpu_buffer->buffer);
2384 diff = ts - cpu_buffer->write_stamp;
2386 /* make sure this diff is calculated here */
2389 /* Did the write stamp get updated already? */
2390 if (likely(ts >= cpu_buffer->write_stamp)) {
2392 if (unlikely(test_time_stamp(delta))) {
2393 int local_clock_stable = 1;
2394 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2395 local_clock_stable = sched_clock_stable;
2397 WARN_ONCE(delta > (1ULL << 59),
2398 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2399 (unsigned long long)delta,
2400 (unsigned long long)ts,
2401 (unsigned long long)cpu_buffer->write_stamp,
2402 local_clock_stable ? "" :
2403 "If you just came from a suspend/resume,\n"
2404 "please switch to the trace global clock:\n"
2405 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2410 event = __rb_reserve_next(cpu_buffer, length, ts,
2411 delta, add_timestamp);
2412 if (unlikely(PTR_ERR(event) == -EAGAIN))
2421 rb_end_commit(cpu_buffer);
2425 #ifdef CONFIG_TRACING
2427 #define TRACE_RECURSIVE_DEPTH 16
2429 /* Keep this code out of the fast path cache */
2430 static noinline void trace_recursive_fail(void)
2432 /* Disable all tracing before we do anything else */
2433 tracing_off_permanent();
2435 printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2436 "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2437 trace_recursion_buffer(),
2438 hardirq_count() >> HARDIRQ_SHIFT,
2439 softirq_count() >> SOFTIRQ_SHIFT,
2445 static inline int trace_recursive_lock(void)
2447 trace_recursion_inc();
2449 if (likely(trace_recursion_buffer() < TRACE_RECURSIVE_DEPTH))
2452 trace_recursive_fail();
2457 static inline void trace_recursive_unlock(void)
2459 WARN_ON_ONCE(!trace_recursion_buffer());
2461 trace_recursion_dec();
2466 #define trace_recursive_lock() (0)
2467 #define trace_recursive_unlock() do { } while (0)
2472 * ring_buffer_lock_reserve - reserve a part of the buffer
2473 * @buffer: the ring buffer to reserve from
2474 * @length: the length of the data to reserve (excluding event header)
2476 * Returns a reseverd event on the ring buffer to copy directly to.
2477 * The user of this interface will need to get the body to write into
2478 * and can use the ring_buffer_event_data() interface.
2480 * The length is the length of the data needed, not the event length
2481 * which also includes the event header.
2483 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2484 * If NULL is returned, then nothing has been allocated or locked.
2486 struct ring_buffer_event *
2487 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2489 struct ring_buffer_per_cpu *cpu_buffer;
2490 struct ring_buffer_event *event;
2493 if (ring_buffer_flags != RB_BUFFERS_ON)
2496 /* If we are tracing schedule, we don't want to recurse */
2497 preempt_disable_notrace();
2499 if (atomic_read(&buffer->record_disabled))
2502 if (trace_recursive_lock())
2505 cpu = raw_smp_processor_id();
2507 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2510 cpu_buffer = buffer->buffers[cpu];
2512 if (atomic_read(&cpu_buffer->record_disabled))
2515 if (length > BUF_MAX_DATA_SIZE)
2518 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2525 trace_recursive_unlock();
2528 preempt_enable_notrace();
2531 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2534 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2535 struct ring_buffer_event *event)
2540 * The event first in the commit queue updates the
2543 if (rb_event_is_commit(cpu_buffer, event)) {
2545 * A commit event that is first on a page
2546 * updates the write timestamp with the page stamp
2548 if (!rb_event_index(event))
2549 cpu_buffer->write_stamp =
2550 cpu_buffer->commit_page->page->time_stamp;
2551 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2552 delta = event->array[0];
2554 delta += event->time_delta;
2555 cpu_buffer->write_stamp += delta;
2557 cpu_buffer->write_stamp += event->time_delta;
2561 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2562 struct ring_buffer_event *event)
2564 local_inc(&cpu_buffer->entries);
2565 rb_update_write_stamp(cpu_buffer, event);
2566 rb_end_commit(cpu_buffer);
2570 * ring_buffer_unlock_commit - commit a reserved
2571 * @buffer: The buffer to commit to
2572 * @event: The event pointer to commit.
2574 * This commits the data to the ring buffer, and releases any locks held.
2576 * Must be paired with ring_buffer_lock_reserve.
2578 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2579 struct ring_buffer_event *event)
2581 struct ring_buffer_per_cpu *cpu_buffer;
2582 int cpu = raw_smp_processor_id();
2584 cpu_buffer = buffer->buffers[cpu];
2586 rb_commit(cpu_buffer, event);
2588 trace_recursive_unlock();
2590 preempt_enable_notrace();
2594 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2596 static inline void rb_event_discard(struct ring_buffer_event *event)
2598 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2599 event = skip_time_extend(event);
2601 /* array[0] holds the actual length for the discarded event */
2602 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2603 event->type_len = RINGBUF_TYPE_PADDING;
2604 /* time delta must be non zero */
2605 if (!event->time_delta)
2606 event->time_delta = 1;
2610 * Decrement the entries to the page that an event is on.
2611 * The event does not even need to exist, only the pointer
2612 * to the page it is on. This may only be called before the commit
2616 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2617 struct ring_buffer_event *event)
2619 unsigned long addr = (unsigned long)event;
2620 struct buffer_page *bpage = cpu_buffer->commit_page;
2621 struct buffer_page *start;
2625 /* Do the likely case first */
2626 if (likely(bpage->page == (void *)addr)) {
2627 local_dec(&bpage->entries);
2632 * Because the commit page may be on the reader page we
2633 * start with the next page and check the end loop there.
2635 rb_inc_page(cpu_buffer, &bpage);
2638 if (bpage->page == (void *)addr) {
2639 local_dec(&bpage->entries);
2642 rb_inc_page(cpu_buffer, &bpage);
2643 } while (bpage != start);
2645 /* commit not part of this buffer?? */
2646 RB_WARN_ON(cpu_buffer, 1);
2650 * ring_buffer_commit_discard - discard an event that has not been committed
2651 * @buffer: the ring buffer
2652 * @event: non committed event to discard
2654 * Sometimes an event that is in the ring buffer needs to be ignored.
2655 * This function lets the user discard an event in the ring buffer
2656 * and then that event will not be read later.
2658 * This function only works if it is called before the the item has been
2659 * committed. It will try to free the event from the ring buffer
2660 * if another event has not been added behind it.
2662 * If another event has been added behind it, it will set the event
2663 * up as discarded, and perform the commit.
2665 * If this function is called, do not call ring_buffer_unlock_commit on
2668 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2669 struct ring_buffer_event *event)
2671 struct ring_buffer_per_cpu *cpu_buffer;
2674 /* The event is discarded regardless */
2675 rb_event_discard(event);
2677 cpu = smp_processor_id();
2678 cpu_buffer = buffer->buffers[cpu];
2681 * This must only be called if the event has not been
2682 * committed yet. Thus we can assume that preemption
2683 * is still disabled.
2685 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2687 rb_decrement_entry(cpu_buffer, event);
2688 if (rb_try_to_discard(cpu_buffer, event))
2692 * The commit is still visible by the reader, so we
2693 * must still update the timestamp.
2695 rb_update_write_stamp(cpu_buffer, event);
2697 rb_end_commit(cpu_buffer);
2699 trace_recursive_unlock();
2701 preempt_enable_notrace();
2704 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2707 * ring_buffer_write - write data to the buffer without reserving
2708 * @buffer: The ring buffer to write to.
2709 * @length: The length of the data being written (excluding the event header)
2710 * @data: The data to write to the buffer.
2712 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2713 * one function. If you already have the data to write to the buffer, it
2714 * may be easier to simply call this function.
2716 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2717 * and not the length of the event which would hold the header.
2719 int ring_buffer_write(struct ring_buffer *buffer,
2720 unsigned long length,
2723 struct ring_buffer_per_cpu *cpu_buffer;
2724 struct ring_buffer_event *event;
2729 if (ring_buffer_flags != RB_BUFFERS_ON)
2732 preempt_disable_notrace();
2734 if (atomic_read(&buffer->record_disabled))
2737 cpu = raw_smp_processor_id();
2739 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2742 cpu_buffer = buffer->buffers[cpu];
2744 if (atomic_read(&cpu_buffer->record_disabled))
2747 if (length > BUF_MAX_DATA_SIZE)
2750 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2754 body = rb_event_data(event);
2756 memcpy(body, data, length);
2758 rb_commit(cpu_buffer, event);
2762 preempt_enable_notrace();
2766 EXPORT_SYMBOL_GPL(ring_buffer_write);
2768 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2770 struct buffer_page *reader = cpu_buffer->reader_page;
2771 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2772 struct buffer_page *commit = cpu_buffer->commit_page;
2774 /* In case of error, head will be NULL */
2775 if (unlikely(!head))
2778 return reader->read == rb_page_commit(reader) &&
2779 (commit == reader ||
2781 head->read == rb_page_commit(commit)));
2785 * ring_buffer_record_disable - stop all writes into the buffer
2786 * @buffer: The ring buffer to stop writes to.
2788 * This prevents all writes to the buffer. Any attempt to write
2789 * to the buffer after this will fail and return NULL.
2791 * The caller should call synchronize_sched() after this.
2793 void ring_buffer_record_disable(struct ring_buffer *buffer)
2795 atomic_inc(&buffer->record_disabled);
2797 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2800 * ring_buffer_record_enable - enable writes to the buffer
2801 * @buffer: The ring buffer to enable writes
2803 * Note, multiple disables will need the same number of enables
2804 * to truly enable the writing (much like preempt_disable).
2806 void ring_buffer_record_enable(struct ring_buffer *buffer)
2808 atomic_dec(&buffer->record_disabled);
2810 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2813 * ring_buffer_record_off - stop all writes into the buffer
2814 * @buffer: The ring buffer to stop writes to.
2816 * This prevents all writes to the buffer. Any attempt to write
2817 * to the buffer after this will fail and return NULL.
2819 * This is different than ring_buffer_record_disable() as
2820 * it works like an on/off switch, where as the disable() verison
2821 * must be paired with a enable().
2823 void ring_buffer_record_off(struct ring_buffer *buffer)
2826 unsigned int new_rd;
2829 rd = atomic_read(&buffer->record_disabled);
2830 new_rd = rd | RB_BUFFER_OFF;
2831 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
2833 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
2836 * ring_buffer_record_on - restart writes into the buffer
2837 * @buffer: The ring buffer to start writes to.
2839 * This enables all writes to the buffer that was disabled by
2840 * ring_buffer_record_off().
2842 * This is different than ring_buffer_record_enable() as
2843 * it works like an on/off switch, where as the enable() verison
2844 * must be paired with a disable().
2846 void ring_buffer_record_on(struct ring_buffer *buffer)
2849 unsigned int new_rd;
2852 rd = atomic_read(&buffer->record_disabled);
2853 new_rd = rd & ~RB_BUFFER_OFF;
2854 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
2856 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
2859 * ring_buffer_record_is_on - return true if the ring buffer can write
2860 * @buffer: The ring buffer to see if write is enabled
2862 * Returns true if the ring buffer is in a state that it accepts writes.
2864 int ring_buffer_record_is_on(struct ring_buffer *buffer)
2866 return !atomic_read(&buffer->record_disabled);
2870 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2871 * @buffer: The ring buffer to stop writes to.
2872 * @cpu: The CPU buffer to stop
2874 * This prevents all writes to the buffer. Any attempt to write
2875 * to the buffer after this will fail and return NULL.
2877 * The caller should call synchronize_sched() after this.
2879 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2881 struct ring_buffer_per_cpu *cpu_buffer;
2883 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2886 cpu_buffer = buffer->buffers[cpu];
2887 atomic_inc(&cpu_buffer->record_disabled);
2889 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2892 * ring_buffer_record_enable_cpu - enable writes to the buffer
2893 * @buffer: The ring buffer to enable writes
2894 * @cpu: The CPU to enable.
2896 * Note, multiple disables will need the same number of enables
2897 * to truly enable the writing (much like preempt_disable).
2899 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2901 struct ring_buffer_per_cpu *cpu_buffer;
2903 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2906 cpu_buffer = buffer->buffers[cpu];
2907 atomic_dec(&cpu_buffer->record_disabled);
2909 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2912 * The total entries in the ring buffer is the running counter
2913 * of entries entered into the ring buffer, minus the sum of
2914 * the entries read from the ring buffer and the number of
2915 * entries that were overwritten.
2917 static inline unsigned long
2918 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
2920 return local_read(&cpu_buffer->entries) -
2921 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
2925 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
2926 * @buffer: The ring buffer
2927 * @cpu: The per CPU buffer to read from.
2929 unsigned long ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
2931 unsigned long flags;
2932 struct ring_buffer_per_cpu *cpu_buffer;
2933 struct buffer_page *bpage;
2936 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2939 cpu_buffer = buffer->buffers[cpu];
2940 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2942 * if the tail is on reader_page, oldest time stamp is on the reader
2945 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
2946 bpage = cpu_buffer->reader_page;
2948 bpage = rb_set_head_page(cpu_buffer);
2949 ret = bpage->page->time_stamp;
2950 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2954 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
2957 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
2958 * @buffer: The ring buffer
2959 * @cpu: The per CPU buffer to read from.
2961 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
2963 struct ring_buffer_per_cpu *cpu_buffer;
2966 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2969 cpu_buffer = buffer->buffers[cpu];
2970 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
2974 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
2977 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2978 * @buffer: The ring buffer
2979 * @cpu: The per CPU buffer to get the entries from.
2981 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2983 struct ring_buffer_per_cpu *cpu_buffer;
2985 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2988 cpu_buffer = buffer->buffers[cpu];
2990 return rb_num_of_entries(cpu_buffer);
2992 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
2995 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
2996 * @buffer: The ring buffer
2997 * @cpu: The per CPU buffer to get the number of overruns from
2999 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3001 struct ring_buffer_per_cpu *cpu_buffer;
3004 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3007 cpu_buffer = buffer->buffers[cpu];
3008 ret = local_read(&cpu_buffer->overrun);
3012 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3015 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
3016 * @buffer: The ring buffer
3017 * @cpu: The per CPU buffer to get the number of overruns from
3020 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3022 struct ring_buffer_per_cpu *cpu_buffer;
3025 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3028 cpu_buffer = buffer->buffers[cpu];
3029 ret = local_read(&cpu_buffer->commit_overrun);
3033 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3036 * ring_buffer_entries - get the number of entries in a buffer
3037 * @buffer: The ring buffer
3039 * Returns the total number of entries in the ring buffer
3042 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3044 struct ring_buffer_per_cpu *cpu_buffer;
3045 unsigned long entries = 0;
3048 /* if you care about this being correct, lock the buffer */
3049 for_each_buffer_cpu(buffer, cpu) {
3050 cpu_buffer = buffer->buffers[cpu];
3051 entries += rb_num_of_entries(cpu_buffer);
3056 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3059 * ring_buffer_overruns - get the number of overruns in buffer
3060 * @buffer: The ring buffer
3062 * Returns the total number of overruns in the ring buffer
3065 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3067 struct ring_buffer_per_cpu *cpu_buffer;
3068 unsigned long overruns = 0;
3071 /* if you care about this being correct, lock the buffer */
3072 for_each_buffer_cpu(buffer, cpu) {
3073 cpu_buffer = buffer->buffers[cpu];
3074 overruns += local_read(&cpu_buffer->overrun);
3079 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3081 static void rb_iter_reset(struct ring_buffer_iter *iter)
3083 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3085 /* Iterator usage is expected to have record disabled */
3086 if (list_empty(&cpu_buffer->reader_page->list)) {
3087 iter->head_page = rb_set_head_page(cpu_buffer);
3088 if (unlikely(!iter->head_page))
3090 iter->head = iter->head_page->read;
3092 iter->head_page = cpu_buffer->reader_page;
3093 iter->head = cpu_buffer->reader_page->read;
3096 iter->read_stamp = cpu_buffer->read_stamp;
3098 iter->read_stamp = iter->head_page->page->time_stamp;
3099 iter->cache_reader_page = cpu_buffer->reader_page;
3100 iter->cache_read = cpu_buffer->read;
3104 * ring_buffer_iter_reset - reset an iterator
3105 * @iter: The iterator to reset
3107 * Resets the iterator, so that it will start from the beginning
3110 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3112 struct ring_buffer_per_cpu *cpu_buffer;
3113 unsigned long flags;
3118 cpu_buffer = iter->cpu_buffer;
3120 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3121 rb_iter_reset(iter);
3122 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3124 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3127 * ring_buffer_iter_empty - check if an iterator has no more to read
3128 * @iter: The iterator to check
3130 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3132 struct ring_buffer_per_cpu *cpu_buffer;
3134 cpu_buffer = iter->cpu_buffer;
3136 return iter->head_page == cpu_buffer->commit_page &&
3137 iter->head == rb_commit_index(cpu_buffer);
3139 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3142 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3143 struct ring_buffer_event *event)
3147 switch (event->type_len) {
3148 case RINGBUF_TYPE_PADDING:
3151 case RINGBUF_TYPE_TIME_EXTEND:
3152 delta = event->array[0];
3154 delta += event->time_delta;
3155 cpu_buffer->read_stamp += delta;
3158 case RINGBUF_TYPE_TIME_STAMP:
3159 /* FIXME: not implemented */
3162 case RINGBUF_TYPE_DATA:
3163 cpu_buffer->read_stamp += event->time_delta;
3173 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3174 struct ring_buffer_event *event)
3178 switch (event->type_len) {
3179 case RINGBUF_TYPE_PADDING:
3182 case RINGBUF_TYPE_TIME_EXTEND:
3183 delta = event->array[0];
3185 delta += event->time_delta;
3186 iter->read_stamp += delta;
3189 case RINGBUF_TYPE_TIME_STAMP:
3190 /* FIXME: not implemented */
3193 case RINGBUF_TYPE_DATA:
3194 iter->read_stamp += event->time_delta;
3203 static struct buffer_page *
3204 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3206 struct buffer_page *reader = NULL;
3207 unsigned long overwrite;
3208 unsigned long flags;
3212 local_irq_save(flags);
3213 arch_spin_lock(&cpu_buffer->lock);
3217 * This should normally only loop twice. But because the
3218 * start of the reader inserts an empty page, it causes
3219 * a case where we will loop three times. There should be no
3220 * reason to loop four times (that I know of).
3222 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3227 reader = cpu_buffer->reader_page;
3229 /* If there's more to read, return this page */
3230 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3233 /* Never should we have an index greater than the size */
3234 if (RB_WARN_ON(cpu_buffer,
3235 cpu_buffer->reader_page->read > rb_page_size(reader)))
3238 /* check if we caught up to the tail */
3240 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3244 * Reset the reader page to size zero.
3246 local_set(&cpu_buffer->reader_page->write, 0);
3247 local_set(&cpu_buffer->reader_page->entries, 0);
3248 local_set(&cpu_buffer->reader_page->page->commit, 0);
3249 cpu_buffer->reader_page->real_end = 0;
3253 * Splice the empty reader page into the list around the head.
3255 reader = rb_set_head_page(cpu_buffer);
3256 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3257 cpu_buffer->reader_page->list.prev = reader->list.prev;
3260 * cpu_buffer->pages just needs to point to the buffer, it
3261 * has no specific buffer page to point to. Lets move it out
3262 * of our way so we don't accidentally swap it.
3264 cpu_buffer->pages = reader->list.prev;
3266 /* The reader page will be pointing to the new head */
3267 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3270 * We want to make sure we read the overruns after we set up our
3271 * pointers to the next object. The writer side does a
3272 * cmpxchg to cross pages which acts as the mb on the writer
3273 * side. Note, the reader will constantly fail the swap
3274 * while the writer is updating the pointers, so this
3275 * guarantees that the overwrite recorded here is the one we
3276 * want to compare with the last_overrun.
3279 overwrite = local_read(&(cpu_buffer->overrun));
3282 * Here's the tricky part.
3284 * We need to move the pointer past the header page.
3285 * But we can only do that if a writer is not currently
3286 * moving it. The page before the header page has the
3287 * flag bit '1' set if it is pointing to the page we want.
3288 * but if the writer is in the process of moving it
3289 * than it will be '2' or already moved '0'.
3292 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3295 * If we did not convert it, then we must try again.
3301 * Yeah! We succeeded in replacing the page.
3303 * Now make the new head point back to the reader page.
3305 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3306 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3308 /* Finally update the reader page to the new head */
3309 cpu_buffer->reader_page = reader;
3310 rb_reset_reader_page(cpu_buffer);
3312 if (overwrite != cpu_buffer->last_overrun) {
3313 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3314 cpu_buffer->last_overrun = overwrite;
3320 arch_spin_unlock(&cpu_buffer->lock);
3321 local_irq_restore(flags);
3326 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3328 struct ring_buffer_event *event;
3329 struct buffer_page *reader;
3332 reader = rb_get_reader_page(cpu_buffer);
3334 /* This function should not be called when buffer is empty */
3335 if (RB_WARN_ON(cpu_buffer, !reader))
3338 event = rb_reader_event(cpu_buffer);
3340 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3343 rb_update_read_stamp(cpu_buffer, event);
3345 length = rb_event_length(event);
3346 cpu_buffer->reader_page->read += length;
3349 static void rb_advance_iter(struct ring_buffer_iter *iter)
3351 struct ring_buffer_per_cpu *cpu_buffer;
3352 struct ring_buffer_event *event;
3355 cpu_buffer = iter->cpu_buffer;
3358 * Check if we are at the end of the buffer.
3360 if (iter->head >= rb_page_size(iter->head_page)) {
3361 /* discarded commits can make the page empty */
3362 if (iter->head_page == cpu_buffer->commit_page)
3368 event = rb_iter_head_event(iter);
3370 length = rb_event_length(event);
3373 * This should not be called to advance the header if we are
3374 * at the tail of the buffer.
3376 if (RB_WARN_ON(cpu_buffer,
3377 (iter->head_page == cpu_buffer->commit_page) &&
3378 (iter->head + length > rb_commit_index(cpu_buffer))))
3381 rb_update_iter_read_stamp(iter, event);
3383 iter->head += length;
3385 /* check for end of page padding */
3386 if ((iter->head >= rb_page_size(iter->head_page)) &&
3387 (iter->head_page != cpu_buffer->commit_page))
3388 rb_advance_iter(iter);
3391 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3393 return cpu_buffer->lost_events;
3396 static struct ring_buffer_event *
3397 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3398 unsigned long *lost_events)
3400 struct ring_buffer_event *event;
3401 struct buffer_page *reader;
3406 * We repeat when a time extend is encountered.
3407 * Since the time extend is always attached to a data event,
3408 * we should never loop more than once.
3409 * (We never hit the following condition more than twice).
3411 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3414 reader = rb_get_reader_page(cpu_buffer);
3418 event = rb_reader_event(cpu_buffer);
3420 switch (event->type_len) {
3421 case RINGBUF_TYPE_PADDING:
3422 if (rb_null_event(event))
3423 RB_WARN_ON(cpu_buffer, 1);
3425 * Because the writer could be discarding every
3426 * event it creates (which would probably be bad)
3427 * if we were to go back to "again" then we may never
3428 * catch up, and will trigger the warn on, or lock
3429 * the box. Return the padding, and we will release
3430 * the current locks, and try again.
3434 case RINGBUF_TYPE_TIME_EXTEND:
3435 /* Internal data, OK to advance */
3436 rb_advance_reader(cpu_buffer);
3439 case RINGBUF_TYPE_TIME_STAMP:
3440 /* FIXME: not implemented */
3441 rb_advance_reader(cpu_buffer);
3444 case RINGBUF_TYPE_DATA:
3446 *ts = cpu_buffer->read_stamp + event->time_delta;
3447 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3448 cpu_buffer->cpu, ts);
3451 *lost_events = rb_lost_events(cpu_buffer);
3460 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3462 static struct ring_buffer_event *
3463 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3465 struct ring_buffer *buffer;
3466 struct ring_buffer_per_cpu *cpu_buffer;
3467 struct ring_buffer_event *event;
3470 cpu_buffer = iter->cpu_buffer;
3471 buffer = cpu_buffer->buffer;
3474 * Check if someone performed a consuming read to
3475 * the buffer. A consuming read invalidates the iterator
3476 * and we need to reset the iterator in this case.
3478 if (unlikely(iter->cache_read != cpu_buffer->read ||
3479 iter->cache_reader_page != cpu_buffer->reader_page))
3480 rb_iter_reset(iter);
3483 if (ring_buffer_iter_empty(iter))
3487 * We repeat when a time extend is encountered.
3488 * Since the time extend is always attached to a data event,
3489 * we should never loop more than once.
3490 * (We never hit the following condition more than twice).
3492 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3495 if (rb_per_cpu_empty(cpu_buffer))
3498 if (iter->head >= local_read(&iter->head_page->page->commit)) {
3503 event = rb_iter_head_event(iter);
3505 switch (event->type_len) {
3506 case RINGBUF_TYPE_PADDING:
3507 if (rb_null_event(event)) {
3511 rb_advance_iter(iter);
3514 case RINGBUF_TYPE_TIME_EXTEND:
3515 /* Internal data, OK to advance */
3516 rb_advance_iter(iter);
3519 case RINGBUF_TYPE_TIME_STAMP:
3520 /* FIXME: not implemented */
3521 rb_advance_iter(iter);
3524 case RINGBUF_TYPE_DATA:
3526 *ts = iter->read_stamp + event->time_delta;
3527 ring_buffer_normalize_time_stamp(buffer,
3528 cpu_buffer->cpu, ts);
3538 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3540 static inline int rb_ok_to_lock(void)
3543 * If an NMI die dumps out the content of the ring buffer
3544 * do not grab locks. We also permanently disable the ring
3545 * buffer too. A one time deal is all you get from reading
3546 * the ring buffer from an NMI.
3548 if (likely(!in_nmi()))
3551 tracing_off_permanent();
3556 * ring_buffer_peek - peek at the next event to be read
3557 * @buffer: The ring buffer to read
3558 * @cpu: The cpu to peak at
3559 * @ts: The timestamp counter of this event.
3560 * @lost_events: a variable to store if events were lost (may be NULL)
3562 * This will return the event that will be read next, but does
3563 * not consume the data.
3565 struct ring_buffer_event *
3566 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3567 unsigned long *lost_events)
3569 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3570 struct ring_buffer_event *event;
3571 unsigned long flags;
3574 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3577 dolock = rb_ok_to_lock();
3579 local_irq_save(flags);
3581 raw_spin_lock(&cpu_buffer->reader_lock);
3582 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3583 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3584 rb_advance_reader(cpu_buffer);
3586 raw_spin_unlock(&cpu_buffer->reader_lock);
3587 local_irq_restore(flags);
3589 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3596 * ring_buffer_iter_peek - peek at the next event to be read
3597 * @iter: The ring buffer iterator
3598 * @ts: The timestamp counter of this event.
3600 * This will return the event that will be read next, but does
3601 * not increment the iterator.
3603 struct ring_buffer_event *
3604 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3606 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3607 struct ring_buffer_event *event;
3608 unsigned long flags;
3611 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3612 event = rb_iter_peek(iter, ts);
3613 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3615 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3622 * ring_buffer_consume - return an event and consume it
3623 * @buffer: The ring buffer to get the next event from
3624 * @cpu: the cpu to read the buffer from
3625 * @ts: a variable to store the timestamp (may be NULL)
3626 * @lost_events: a variable to store if events were lost (may be NULL)
3628 * Returns the next event in the ring buffer, and that event is consumed.
3629 * Meaning, that sequential reads will keep returning a different event,
3630 * and eventually empty the ring buffer if the producer is slower.
3632 struct ring_buffer_event *
3633 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3634 unsigned long *lost_events)
3636 struct ring_buffer_per_cpu *cpu_buffer;
3637 struct ring_buffer_event *event = NULL;
3638 unsigned long flags;
3641 dolock = rb_ok_to_lock();
3644 /* might be called in atomic */
3647 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3650 cpu_buffer = buffer->buffers[cpu];
3651 local_irq_save(flags);
3653 raw_spin_lock(&cpu_buffer->reader_lock);
3655 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3657 cpu_buffer->lost_events = 0;
3658 rb_advance_reader(cpu_buffer);
3662 raw_spin_unlock(&cpu_buffer->reader_lock);
3663 local_irq_restore(flags);
3668 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3673 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3676 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3677 * @buffer: The ring buffer to read from
3678 * @cpu: The cpu buffer to iterate over
3680 * This performs the initial preparations necessary to iterate
3681 * through the buffer. Memory is allocated, buffer recording
3682 * is disabled, and the iterator pointer is returned to the caller.
3684 * Disabling buffer recordng prevents the reading from being
3685 * corrupted. This is not a consuming read, so a producer is not
3688 * After a sequence of ring_buffer_read_prepare calls, the user is
3689 * expected to make at least one call to ring_buffer_prepare_sync.
3690 * Afterwards, ring_buffer_read_start is invoked to get things going
3693 * This overall must be paired with ring_buffer_finish.
3695 struct ring_buffer_iter *
3696 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3698 struct ring_buffer_per_cpu *cpu_buffer;
3699 struct ring_buffer_iter *iter;
3701 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3704 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3708 cpu_buffer = buffer->buffers[cpu];
3710 iter->cpu_buffer = cpu_buffer;
3712 atomic_inc(&buffer->resize_disabled);
3713 atomic_inc(&cpu_buffer->record_disabled);
3717 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3720 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3722 * All previously invoked ring_buffer_read_prepare calls to prepare
3723 * iterators will be synchronized. Afterwards, read_buffer_read_start
3724 * calls on those iterators are allowed.
3727 ring_buffer_read_prepare_sync(void)
3729 synchronize_sched();
3731 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
3734 * ring_buffer_read_start - start a non consuming read of the buffer
3735 * @iter: The iterator returned by ring_buffer_read_prepare
3737 * This finalizes the startup of an iteration through the buffer.
3738 * The iterator comes from a call to ring_buffer_read_prepare and
3739 * an intervening ring_buffer_read_prepare_sync must have been
3742 * Must be paired with ring_buffer_finish.
3745 ring_buffer_read_start(struct ring_buffer_iter *iter)
3747 struct ring_buffer_per_cpu *cpu_buffer;
3748 unsigned long flags;
3753 cpu_buffer = iter->cpu_buffer;
3755 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3756 arch_spin_lock(&cpu_buffer->lock);
3757 rb_iter_reset(iter);
3758 arch_spin_unlock(&cpu_buffer->lock);
3759 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3761 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3764 * ring_buffer_finish - finish reading the iterator of the buffer
3765 * @iter: The iterator retrieved by ring_buffer_start
3767 * This re-enables the recording to the buffer, and frees the
3771 ring_buffer_read_finish(struct ring_buffer_iter *iter)
3773 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3776 * Ring buffer is disabled from recording, here's a good place
3777 * to check the integrity of the ring buffer.
3779 rb_check_pages(cpu_buffer);
3781 atomic_dec(&cpu_buffer->record_disabled);
3782 atomic_dec(&cpu_buffer->buffer->resize_disabled);
3785 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3788 * ring_buffer_read - read the next item in the ring buffer by the iterator
3789 * @iter: The ring buffer iterator
3790 * @ts: The time stamp of the event read.
3792 * This reads the next event in the ring buffer and increments the iterator.
3794 struct ring_buffer_event *
3795 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3797 struct ring_buffer_event *event;
3798 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3799 unsigned long flags;
3801 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3803 event = rb_iter_peek(iter, ts);
3807 if (event->type_len == RINGBUF_TYPE_PADDING)
3810 rb_advance_iter(iter);
3812 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3816 EXPORT_SYMBOL_GPL(ring_buffer_read);
3819 * ring_buffer_size - return the size of the ring buffer (in bytes)
3820 * @buffer: The ring buffer.
3822 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
3825 * Earlier, this method returned
3826 * BUF_PAGE_SIZE * buffer->nr_pages
3827 * Since the nr_pages field is now removed, we have converted this to
3828 * return the per cpu buffer value.
3830 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3833 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
3835 EXPORT_SYMBOL_GPL(ring_buffer_size);
3838 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3840 rb_head_page_deactivate(cpu_buffer);
3842 cpu_buffer->head_page
3843 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3844 local_set(&cpu_buffer->head_page->write, 0);
3845 local_set(&cpu_buffer->head_page->entries, 0);
3846 local_set(&cpu_buffer->head_page->page->commit, 0);
3848 cpu_buffer->head_page->read = 0;
3850 cpu_buffer->tail_page = cpu_buffer->head_page;
3851 cpu_buffer->commit_page = cpu_buffer->head_page;
3853 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3854 INIT_LIST_HEAD(&cpu_buffer->new_pages);
3855 local_set(&cpu_buffer->reader_page->write, 0);
3856 local_set(&cpu_buffer->reader_page->entries, 0);
3857 local_set(&cpu_buffer->reader_page->page->commit, 0);
3858 cpu_buffer->reader_page->read = 0;
3860 local_set(&cpu_buffer->commit_overrun, 0);
3861 local_set(&cpu_buffer->entries_bytes, 0);
3862 local_set(&cpu_buffer->overrun, 0);
3863 local_set(&cpu_buffer->entries, 0);
3864 local_set(&cpu_buffer->committing, 0);
3865 local_set(&cpu_buffer->commits, 0);
3866 cpu_buffer->read = 0;
3867 cpu_buffer->read_bytes = 0;
3869 cpu_buffer->write_stamp = 0;
3870 cpu_buffer->read_stamp = 0;
3872 cpu_buffer->lost_events = 0;
3873 cpu_buffer->last_overrun = 0;
3875 rb_head_page_activate(cpu_buffer);
3879 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3880 * @buffer: The ring buffer to reset a per cpu buffer of
3881 * @cpu: The CPU buffer to be reset
3883 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3885 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3886 unsigned long flags;
3888 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3891 atomic_inc(&buffer->resize_disabled);
3892 atomic_inc(&cpu_buffer->record_disabled);
3894 /* Make sure all commits have finished */
3895 synchronize_sched();
3897 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3899 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3902 arch_spin_lock(&cpu_buffer->lock);
3904 rb_reset_cpu(cpu_buffer);
3906 arch_spin_unlock(&cpu_buffer->lock);
3909 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3911 atomic_dec(&cpu_buffer->record_disabled);
3912 atomic_dec(&buffer->resize_disabled);
3914 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3917 * ring_buffer_reset - reset a ring buffer
3918 * @buffer: The ring buffer to reset all cpu buffers
3920 void ring_buffer_reset(struct ring_buffer *buffer)
3924 for_each_buffer_cpu(buffer, cpu)
3925 ring_buffer_reset_cpu(buffer, cpu);
3927 EXPORT_SYMBOL_GPL(ring_buffer_reset);
3930 * rind_buffer_empty - is the ring buffer empty?
3931 * @buffer: The ring buffer to test
3933 int ring_buffer_empty(struct ring_buffer *buffer)
3935 struct ring_buffer_per_cpu *cpu_buffer;
3936 unsigned long flags;
3941 dolock = rb_ok_to_lock();
3943 /* yes this is racy, but if you don't like the race, lock the buffer */
3944 for_each_buffer_cpu(buffer, cpu) {
3945 cpu_buffer = buffer->buffers[cpu];
3946 local_irq_save(flags);
3948 raw_spin_lock(&cpu_buffer->reader_lock);
3949 ret = rb_per_cpu_empty(cpu_buffer);
3951 raw_spin_unlock(&cpu_buffer->reader_lock);
3952 local_irq_restore(flags);
3960 EXPORT_SYMBOL_GPL(ring_buffer_empty);
3963 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3964 * @buffer: The ring buffer
3965 * @cpu: The CPU buffer to test
3967 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3969 struct ring_buffer_per_cpu *cpu_buffer;
3970 unsigned long flags;
3974 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3977 dolock = rb_ok_to_lock();
3979 cpu_buffer = buffer->buffers[cpu];
3980 local_irq_save(flags);
3982 raw_spin_lock(&cpu_buffer->reader_lock);
3983 ret = rb_per_cpu_empty(cpu_buffer);
3985 raw_spin_unlock(&cpu_buffer->reader_lock);
3986 local_irq_restore(flags);
3990 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
3992 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
3994 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
3995 * @buffer_a: One buffer to swap with
3996 * @buffer_b: The other buffer to swap with
3998 * This function is useful for tracers that want to take a "snapshot"
3999 * of a CPU buffer and has another back up buffer lying around.
4000 * it is expected that the tracer handles the cpu buffer not being
4001 * used at the moment.
4003 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4004 struct ring_buffer *buffer_b, int cpu)
4006 struct ring_buffer_per_cpu *cpu_buffer_a;
4007 struct ring_buffer_per_cpu *cpu_buffer_b;
4010 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4011 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4014 cpu_buffer_a = buffer_a->buffers[cpu];
4015 cpu_buffer_b = buffer_b->buffers[cpu];
4017 /* At least make sure the two buffers are somewhat the same */
4018 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4023 if (ring_buffer_flags != RB_BUFFERS_ON)
4026 if (atomic_read(&buffer_a->record_disabled))
4029 if (atomic_read(&buffer_b->record_disabled))
4032 if (atomic_read(&cpu_buffer_a->record_disabled))
4035 if (atomic_read(&cpu_buffer_b->record_disabled))
4039 * We can't do a synchronize_sched here because this
4040 * function can be called in atomic context.
4041 * Normally this will be called from the same CPU as cpu.
4042 * If not it's up to the caller to protect this.
4044 atomic_inc(&cpu_buffer_a->record_disabled);
4045 atomic_inc(&cpu_buffer_b->record_disabled);
4048 if (local_read(&cpu_buffer_a->committing))
4050 if (local_read(&cpu_buffer_b->committing))
4053 buffer_a->buffers[cpu] = cpu_buffer_b;
4054 buffer_b->buffers[cpu] = cpu_buffer_a;
4056 cpu_buffer_b->buffer = buffer_a;
4057 cpu_buffer_a->buffer = buffer_b;
4062 atomic_dec(&cpu_buffer_a->record_disabled);
4063 atomic_dec(&cpu_buffer_b->record_disabled);
4067 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4068 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4071 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4072 * @buffer: the buffer to allocate for.
4074 * This function is used in conjunction with ring_buffer_read_page.
4075 * When reading a full page from the ring buffer, these functions
4076 * can be used to speed up the process. The calling function should
4077 * allocate a few pages first with this function. Then when it
4078 * needs to get pages from the ring buffer, it passes the result
4079 * of this function into ring_buffer_read_page, which will swap
4080 * the page that was allocated, with the read page of the buffer.
4083 * The page allocated, or NULL on error.
4085 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4087 struct buffer_data_page *bpage;
4090 page = alloc_pages_node(cpu_to_node(cpu),
4091 GFP_KERNEL | __GFP_NORETRY, 0);
4095 bpage = page_address(page);
4097 rb_init_page(bpage);
4101 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4104 * ring_buffer_free_read_page - free an allocated read page
4105 * @buffer: the buffer the page was allocate for
4106 * @data: the page to free
4108 * Free a page allocated from ring_buffer_alloc_read_page.
4110 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4112 free_page((unsigned long)data);
4114 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4117 * ring_buffer_read_page - extract a page from the ring buffer
4118 * @buffer: buffer to extract from
4119 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4120 * @len: amount to extract
4121 * @cpu: the cpu of the buffer to extract
4122 * @full: should the extraction only happen when the page is full.
4124 * This function will pull out a page from the ring buffer and consume it.
4125 * @data_page must be the address of the variable that was returned
4126 * from ring_buffer_alloc_read_page. This is because the page might be used
4127 * to swap with a page in the ring buffer.
4130 * rpage = ring_buffer_alloc_read_page(buffer);
4133 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4135 * process_page(rpage, ret);
4137 * When @full is set, the function will not return true unless
4138 * the writer is off the reader page.
4140 * Note: it is up to the calling functions to handle sleeps and wakeups.
4141 * The ring buffer can be used anywhere in the kernel and can not
4142 * blindly call wake_up. The layer that uses the ring buffer must be
4143 * responsible for that.
4146 * >=0 if data has been transferred, returns the offset of consumed data.
4147 * <0 if no data has been transferred.
4149 int ring_buffer_read_page(struct ring_buffer *buffer,
4150 void **data_page, size_t len, int cpu, int full)
4152 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4153 struct ring_buffer_event *event;
4154 struct buffer_data_page *bpage;
4155 struct buffer_page *reader;
4156 unsigned long missed_events;
4157 unsigned long flags;
4158 unsigned int commit;
4163 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4167 * If len is not big enough to hold the page header, then
4168 * we can not copy anything.
4170 if (len <= BUF_PAGE_HDR_SIZE)
4173 len -= BUF_PAGE_HDR_SIZE;
4182 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4184 reader = rb_get_reader_page(cpu_buffer);
4188 event = rb_reader_event(cpu_buffer);
4190 read = reader->read;
4191 commit = rb_page_commit(reader);
4193 /* Check if any events were dropped */
4194 missed_events = cpu_buffer->lost_events;
4197 * If this page has been partially read or
4198 * if len is not big enough to read the rest of the page or
4199 * a writer is still on the page, then
4200 * we must copy the data from the page to the buffer.
4201 * Otherwise, we can simply swap the page with the one passed in.
4203 if (read || (len < (commit - read)) ||
4204 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4205 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4206 unsigned int rpos = read;
4207 unsigned int pos = 0;
4213 if (len > (commit - read))
4214 len = (commit - read);
4216 /* Always keep the time extend and data together */
4217 size = rb_event_ts_length(event);
4222 /* save the current timestamp, since the user will need it */
4223 save_timestamp = cpu_buffer->read_stamp;
4225 /* Need to copy one event at a time */
4227 /* We need the size of one event, because
4228 * rb_advance_reader only advances by one event,
4229 * whereas rb_event_ts_length may include the size of
4230 * one or two events.
4231 * We have already ensured there's enough space if this
4232 * is a time extend. */
4233 size = rb_event_length(event);
4234 memcpy(bpage->data + pos, rpage->data + rpos, size);
4238 rb_advance_reader(cpu_buffer);
4239 rpos = reader->read;
4245 event = rb_reader_event(cpu_buffer);
4246 /* Always keep the time extend and data together */
4247 size = rb_event_ts_length(event);
4248 } while (len >= size);
4251 local_set(&bpage->commit, pos);
4252 bpage->time_stamp = save_timestamp;
4254 /* we copied everything to the beginning */
4257 /* update the entry counter */
4258 cpu_buffer->read += rb_page_entries(reader);
4259 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4261 /* swap the pages */
4262 rb_init_page(bpage);
4263 bpage = reader->page;
4264 reader->page = *data_page;
4265 local_set(&reader->write, 0);
4266 local_set(&reader->entries, 0);
4271 * Use the real_end for the data size,
4272 * This gives us a chance to store the lost events
4275 if (reader->real_end)
4276 local_set(&bpage->commit, reader->real_end);
4280 cpu_buffer->lost_events = 0;
4282 commit = local_read(&bpage->commit);
4284 * Set a flag in the commit field if we lost events
4286 if (missed_events) {
4287 /* If there is room at the end of the page to save the
4288 * missed events, then record it there.
4290 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4291 memcpy(&bpage->data[commit], &missed_events,
4292 sizeof(missed_events));
4293 local_add(RB_MISSED_STORED, &bpage->commit);
4294 commit += sizeof(missed_events);
4296 local_add(RB_MISSED_EVENTS, &bpage->commit);
4300 * This page may be off to user land. Zero it out here.
4302 if (commit < BUF_PAGE_SIZE)
4303 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4306 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4311 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4313 #ifdef CONFIG_HOTPLUG_CPU
4314 static int rb_cpu_notify(struct notifier_block *self,
4315 unsigned long action, void *hcpu)
4317 struct ring_buffer *buffer =
4318 container_of(self, struct ring_buffer, cpu_notify);
4319 long cpu = (long)hcpu;
4320 int cpu_i, nr_pages_same;
4321 unsigned int nr_pages;
4324 case CPU_UP_PREPARE:
4325 case CPU_UP_PREPARE_FROZEN:
4326 if (cpumask_test_cpu(cpu, buffer->cpumask))
4331 /* check if all cpu sizes are same */
4332 for_each_buffer_cpu(buffer, cpu_i) {
4333 /* fill in the size from first enabled cpu */
4335 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4336 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4341 /* allocate minimum pages, user can later expand it */
4344 buffer->buffers[cpu] =
4345 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4346 if (!buffer->buffers[cpu]) {
4347 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4352 cpumask_set_cpu(cpu, buffer->cpumask);
4354 case CPU_DOWN_PREPARE:
4355 case CPU_DOWN_PREPARE_FROZEN:
4358 * If we were to free the buffer, then the user would
4359 * lose any trace that was in the buffer.