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 * Increment overrun to account for the lost events.
1352 local_add(page_entries, &cpu_buffer->overrun);
1353 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1357 * We have already removed references to this list item, just
1358 * free up the buffer_page and its page
1360 free_buffer_page(to_remove_page);
1363 } while (to_remove_page != last_page);
1365 RB_WARN_ON(cpu_buffer, nr_removed);
1367 return nr_removed == 0;
1371 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer)
1373 struct list_head *pages = &cpu_buffer->new_pages;
1374 int retries, success;
1376 raw_spin_lock_irq(&cpu_buffer->reader_lock);
1378 * We are holding the reader lock, so the reader page won't be swapped
1379 * in the ring buffer. Now we are racing with the writer trying to
1380 * move head page and the tail page.
1381 * We are going to adapt the reader page update process where:
1382 * 1. We first splice the start and end of list of new pages between
1383 * the head page and its previous page.
1384 * 2. We cmpxchg the prev_page->next to point from head page to the
1385 * start of new pages list.
1386 * 3. Finally, we update the head->prev to the end of new list.
1388 * We will try this process 10 times, to make sure that we don't keep
1394 struct list_head *head_page, *prev_page, *r;
1395 struct list_head *last_page, *first_page;
1396 struct list_head *head_page_with_bit;
1398 head_page = &rb_set_head_page(cpu_buffer)->list;
1401 prev_page = head_page->prev;
1403 first_page = pages->next;
1404 last_page = pages->prev;
1406 head_page_with_bit = (struct list_head *)
1407 ((unsigned long)head_page | RB_PAGE_HEAD);
1409 last_page->next = head_page_with_bit;
1410 first_page->prev = prev_page;
1412 r = cmpxchg(&prev_page->next, head_page_with_bit, first_page);
1414 if (r == head_page_with_bit) {
1416 * yay, we replaced the page pointer to our new list,
1417 * now, we just have to update to head page's prev
1418 * pointer to point to end of list
1420 head_page->prev = last_page;
1427 INIT_LIST_HEAD(pages);
1429 * If we weren't successful in adding in new pages, warn and stop
1432 RB_WARN_ON(cpu_buffer, !success);
1433 raw_spin_unlock_irq(&cpu_buffer->reader_lock);
1435 /* free pages if they weren't inserted */
1437 struct buffer_page *bpage, *tmp;
1438 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1440 list_del_init(&bpage->list);
1441 free_buffer_page(bpage);
1447 static void rb_update_pages(struct ring_buffer_per_cpu *cpu_buffer)
1451 if (cpu_buffer->nr_pages_to_update > 0)
1452 success = rb_insert_pages(cpu_buffer);
1454 success = rb_remove_pages(cpu_buffer,
1455 -cpu_buffer->nr_pages_to_update);
1458 cpu_buffer->nr_pages += cpu_buffer->nr_pages_to_update;
1461 static void update_pages_handler(struct work_struct *work)
1463 struct ring_buffer_per_cpu *cpu_buffer = container_of(work,
1464 struct ring_buffer_per_cpu, update_pages_work);
1465 rb_update_pages(cpu_buffer);
1466 complete(&cpu_buffer->update_done);
1470 * ring_buffer_resize - resize the ring buffer
1471 * @buffer: the buffer to resize.
1472 * @size: the new size.
1474 * Minimum size is 2 * BUF_PAGE_SIZE.
1476 * Returns 0 on success and < 0 on failure.
1478 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size,
1481 struct ring_buffer_per_cpu *cpu_buffer;
1486 * Always succeed at resizing a non-existent buffer:
1491 /* Make sure the requested buffer exists */
1492 if (cpu_id != RING_BUFFER_ALL_CPUS &&
1493 !cpumask_test_cpu(cpu_id, buffer->cpumask))
1496 size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1497 size *= BUF_PAGE_SIZE;
1499 /* we need a minimum of two pages */
1500 if (size < BUF_PAGE_SIZE * 2)
1501 size = BUF_PAGE_SIZE * 2;
1503 nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
1506 * Don't succeed if resizing is disabled, as a reader might be
1507 * manipulating the ring buffer and is expecting a sane state while
1510 if (atomic_read(&buffer->resize_disabled))
1513 /* prevent another thread from changing buffer sizes */
1514 mutex_lock(&buffer->mutex);
1516 if (cpu_id == RING_BUFFER_ALL_CPUS) {
1517 /* calculate the pages to update */
1518 for_each_buffer_cpu(buffer, cpu) {
1519 cpu_buffer = buffer->buffers[cpu];
1521 cpu_buffer->nr_pages_to_update = nr_pages -
1522 cpu_buffer->nr_pages;
1524 * nothing more to do for removing pages or no update
1526 if (cpu_buffer->nr_pages_to_update <= 0)
1529 * to add pages, make sure all new pages can be
1530 * allocated without receiving ENOMEM
1532 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1533 if (__rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1534 &cpu_buffer->new_pages, cpu)) {
1535 /* not enough memory for new pages */
1543 * Fire off all the required work handlers
1544 * We can't schedule on offline CPUs, but it's not necessary
1545 * since we can change their buffer sizes without any race.
1547 for_each_buffer_cpu(buffer, cpu) {
1548 cpu_buffer = buffer->buffers[cpu];
1549 if (!cpu_buffer->nr_pages_to_update)
1552 if (cpu_online(cpu))
1553 schedule_work_on(cpu,
1554 &cpu_buffer->update_pages_work);
1556 rb_update_pages(cpu_buffer);
1559 /* wait for all the updates to complete */
1560 for_each_buffer_cpu(buffer, cpu) {
1561 cpu_buffer = buffer->buffers[cpu];
1562 if (!cpu_buffer->nr_pages_to_update)
1565 if (cpu_online(cpu))
1566 wait_for_completion(&cpu_buffer->update_done);
1567 cpu_buffer->nr_pages_to_update = 0;
1572 /* Make sure this CPU has been intitialized */
1573 if (!cpumask_test_cpu(cpu_id, buffer->cpumask))
1576 cpu_buffer = buffer->buffers[cpu_id];
1578 if (nr_pages == cpu_buffer->nr_pages)
1581 cpu_buffer->nr_pages_to_update = nr_pages -
1582 cpu_buffer->nr_pages;
1584 INIT_LIST_HEAD(&cpu_buffer->new_pages);
1585 if (cpu_buffer->nr_pages_to_update > 0 &&
1586 __rb_allocate_pages(cpu_buffer->nr_pages_to_update,
1587 &cpu_buffer->new_pages, cpu_id)) {
1594 if (cpu_online(cpu_id)) {
1595 schedule_work_on(cpu_id,
1596 &cpu_buffer->update_pages_work);
1597 wait_for_completion(&cpu_buffer->update_done);
1599 rb_update_pages(cpu_buffer);
1601 cpu_buffer->nr_pages_to_update = 0;
1607 * The ring buffer resize can happen with the ring buffer
1608 * enabled, so that the update disturbs the tracing as little
1609 * as possible. But if the buffer is disabled, we do not need
1610 * to worry about that, and we can take the time to verify
1611 * that the buffer is not corrupt.
1613 if (atomic_read(&buffer->record_disabled)) {
1614 atomic_inc(&buffer->record_disabled);
1616 * Even though the buffer was disabled, we must make sure
1617 * that it is truly disabled before calling rb_check_pages.
1618 * There could have been a race between checking
1619 * record_disable and incrementing it.
1621 synchronize_sched();
1622 for_each_buffer_cpu(buffer, cpu) {
1623 cpu_buffer = buffer->buffers[cpu];
1624 rb_check_pages(cpu_buffer);
1626 atomic_dec(&buffer->record_disabled);
1629 mutex_unlock(&buffer->mutex);
1633 for_each_buffer_cpu(buffer, cpu) {
1634 struct buffer_page *bpage, *tmp;
1636 cpu_buffer = buffer->buffers[cpu];
1637 cpu_buffer->nr_pages_to_update = 0;
1639 if (list_empty(&cpu_buffer->new_pages))
1642 list_for_each_entry_safe(bpage, tmp, &cpu_buffer->new_pages,
1644 list_del_init(&bpage->list);
1645 free_buffer_page(bpage);
1648 mutex_unlock(&buffer->mutex);
1651 EXPORT_SYMBOL_GPL(ring_buffer_resize);
1653 void ring_buffer_change_overwrite(struct ring_buffer *buffer, int val)
1655 mutex_lock(&buffer->mutex);
1657 buffer->flags |= RB_FL_OVERWRITE;
1659 buffer->flags &= ~RB_FL_OVERWRITE;
1660 mutex_unlock(&buffer->mutex);
1662 EXPORT_SYMBOL_GPL(ring_buffer_change_overwrite);
1664 static inline void *
1665 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
1667 return bpage->data + index;
1670 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
1672 return bpage->page->data + index;
1675 static inline struct ring_buffer_event *
1676 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
1678 return __rb_page_index(cpu_buffer->reader_page,
1679 cpu_buffer->reader_page->read);
1682 static inline struct ring_buffer_event *
1683 rb_iter_head_event(struct ring_buffer_iter *iter)
1685 return __rb_page_index(iter->head_page, iter->head);
1688 static inline unsigned rb_page_commit(struct buffer_page *bpage)
1690 return local_read(&bpage->page->commit);
1693 /* Size is determined by what has been committed */
1694 static inline unsigned rb_page_size(struct buffer_page *bpage)
1696 return rb_page_commit(bpage);
1699 static inline unsigned
1700 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
1702 return rb_page_commit(cpu_buffer->commit_page);
1705 static inline unsigned
1706 rb_event_index(struct ring_buffer_event *event)
1708 unsigned long addr = (unsigned long)event;
1710 return (addr & ~PAGE_MASK) - BUF_PAGE_HDR_SIZE;
1714 rb_event_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
1715 struct ring_buffer_event *event)
1717 unsigned long addr = (unsigned long)event;
1718 unsigned long index;
1720 index = rb_event_index(event);
1723 return cpu_buffer->commit_page->page == (void *)addr &&
1724 rb_commit_index(cpu_buffer) == index;
1728 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1730 unsigned long max_count;
1733 * We only race with interrupts and NMIs on this CPU.
1734 * If we own the commit event, then we can commit
1735 * all others that interrupted us, since the interruptions
1736 * are in stack format (they finish before they come
1737 * back to us). This allows us to do a simple loop to
1738 * assign the commit to the tail.
1741 max_count = cpu_buffer->nr_pages * 100;
1743 while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1744 if (RB_WARN_ON(cpu_buffer, !(--max_count)))
1746 if (RB_WARN_ON(cpu_buffer,
1747 rb_is_reader_page(cpu_buffer->tail_page)))
1749 local_set(&cpu_buffer->commit_page->page->commit,
1750 rb_page_write(cpu_buffer->commit_page));
1751 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1752 cpu_buffer->write_stamp =
1753 cpu_buffer->commit_page->page->time_stamp;
1754 /* add barrier to keep gcc from optimizing too much */
1757 while (rb_commit_index(cpu_buffer) !=
1758 rb_page_write(cpu_buffer->commit_page)) {
1760 local_set(&cpu_buffer->commit_page->page->commit,
1761 rb_page_write(cpu_buffer->commit_page));
1762 RB_WARN_ON(cpu_buffer,
1763 local_read(&cpu_buffer->commit_page->page->commit) &
1768 /* again, keep gcc from optimizing */
1772 * If an interrupt came in just after the first while loop
1773 * and pushed the tail page forward, we will be left with
1774 * a dangling commit that will never go forward.
1776 if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1780 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1782 cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1783 cpu_buffer->reader_page->read = 0;
1786 static void rb_inc_iter(struct ring_buffer_iter *iter)
1788 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1791 * The iterator could be on the reader page (it starts there).
1792 * But the head could have moved, since the reader was
1793 * found. Check for this case and assign the iterator
1794 * to the head page instead of next.
1796 if (iter->head_page == cpu_buffer->reader_page)
1797 iter->head_page = rb_set_head_page(cpu_buffer);
1799 rb_inc_page(cpu_buffer, &iter->head_page);
1801 iter->read_stamp = iter->head_page->page->time_stamp;
1805 /* Slow path, do not inline */
1806 static noinline struct ring_buffer_event *
1807 rb_add_time_stamp(struct ring_buffer_event *event, u64 delta)
1809 event->type_len = RINGBUF_TYPE_TIME_EXTEND;
1811 /* Not the first event on the page? */
1812 if (rb_event_index(event)) {
1813 event->time_delta = delta & TS_MASK;
1814 event->array[0] = delta >> TS_SHIFT;
1816 /* nope, just zero it */
1817 event->time_delta = 0;
1818 event->array[0] = 0;
1821 return skip_time_extend(event);
1825 * ring_buffer_update_event - update event type and data
1826 * @event: the even to update
1827 * @type: the type of event
1828 * @length: the size of the event field in the ring buffer
1830 * Update the type and data fields of the event. The length
1831 * is the actual size that is written to the ring buffer,
1832 * and with this, we can determine what to place into the
1836 rb_update_event(struct ring_buffer_per_cpu *cpu_buffer,
1837 struct ring_buffer_event *event, unsigned length,
1838 int add_timestamp, u64 delta)
1840 /* Only a commit updates the timestamp */
1841 if (unlikely(!rb_event_is_commit(cpu_buffer, event)))
1845 * If we need to add a timestamp, then we
1846 * add it to the start of the resevered space.
1848 if (unlikely(add_timestamp)) {
1849 event = rb_add_time_stamp(event, delta);
1850 length -= RB_LEN_TIME_EXTEND;
1854 event->time_delta = delta;
1855 length -= RB_EVNT_HDR_SIZE;
1856 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT) {
1857 event->type_len = 0;
1858 event->array[0] = length;
1860 event->type_len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1864 * rb_handle_head_page - writer hit the head page
1866 * Returns: +1 to retry page
1871 rb_handle_head_page(struct ring_buffer_per_cpu *cpu_buffer,
1872 struct buffer_page *tail_page,
1873 struct buffer_page *next_page)
1875 struct buffer_page *new_head;
1880 entries = rb_page_entries(next_page);
1883 * The hard part is here. We need to move the head
1884 * forward, and protect against both readers on
1885 * other CPUs and writers coming in via interrupts.
1887 type = rb_head_page_set_update(cpu_buffer, next_page, tail_page,
1891 * type can be one of four:
1892 * NORMAL - an interrupt already moved it for us
1893 * HEAD - we are the first to get here.
1894 * UPDATE - we are the interrupt interrupting
1896 * MOVED - a reader on another CPU moved the next
1897 * pointer to its reader page. Give up
1904 * We changed the head to UPDATE, thus
1905 * it is our responsibility to update
1908 local_add(entries, &cpu_buffer->overrun);
1909 local_sub(BUF_PAGE_SIZE, &cpu_buffer->entries_bytes);
1912 * The entries will be zeroed out when we move the
1916 /* still more to do */
1919 case RB_PAGE_UPDATE:
1921 * This is an interrupt that interrupt the
1922 * previous update. Still more to do.
1925 case RB_PAGE_NORMAL:
1927 * An interrupt came in before the update
1928 * and processed this for us.
1929 * Nothing left to do.
1934 * The reader is on another CPU and just did
1935 * a swap with our next_page.
1940 RB_WARN_ON(cpu_buffer, 1); /* WTF??? */
1945 * Now that we are here, the old head pointer is
1946 * set to UPDATE. This will keep the reader from
1947 * swapping the head page with the reader page.
1948 * The reader (on another CPU) will spin till
1951 * We just need to protect against interrupts
1952 * doing the job. We will set the next pointer
1953 * to HEAD. After that, we set the old pointer
1954 * to NORMAL, but only if it was HEAD before.
1955 * otherwise we are an interrupt, and only
1956 * want the outer most commit to reset it.
1958 new_head = next_page;
1959 rb_inc_page(cpu_buffer, &new_head);
1961 ret = rb_head_page_set_head(cpu_buffer, new_head, next_page,
1965 * Valid returns are:
1966 * HEAD - an interrupt came in and already set it.
1967 * NORMAL - One of two things:
1968 * 1) We really set it.
1969 * 2) A bunch of interrupts came in and moved
1970 * the page forward again.
1974 case RB_PAGE_NORMAL:
1978 RB_WARN_ON(cpu_buffer, 1);
1983 * It is possible that an interrupt came in,
1984 * set the head up, then more interrupts came in
1985 * and moved it again. When we get back here,
1986 * the page would have been set to NORMAL but we
1987 * just set it back to HEAD.
1989 * How do you detect this? Well, if that happened
1990 * the tail page would have moved.
1992 if (ret == RB_PAGE_NORMAL) {
1994 * If the tail had moved passed next, then we need
1995 * to reset the pointer.
1997 if (cpu_buffer->tail_page != tail_page &&
1998 cpu_buffer->tail_page != next_page)
1999 rb_head_page_set_normal(cpu_buffer, new_head,
2005 * If this was the outer most commit (the one that
2006 * changed the original pointer from HEAD to UPDATE),
2007 * then it is up to us to reset it to NORMAL.
2009 if (type == RB_PAGE_HEAD) {
2010 ret = rb_head_page_set_normal(cpu_buffer, next_page,
2013 if (RB_WARN_ON(cpu_buffer,
2014 ret != RB_PAGE_UPDATE))
2021 static unsigned rb_calculate_event_length(unsigned length)
2023 struct ring_buffer_event event; /* Used only for sizeof array */
2025 /* zero length can cause confusions */
2029 if (length > RB_MAX_SMALL_DATA || RB_FORCE_8BYTE_ALIGNMENT)
2030 length += sizeof(event.array[0]);
2032 length += RB_EVNT_HDR_SIZE;
2033 length = ALIGN(length, RB_ARCH_ALIGNMENT);
2039 rb_reset_tail(struct ring_buffer_per_cpu *cpu_buffer,
2040 struct buffer_page *tail_page,
2041 unsigned long tail, unsigned long length)
2043 struct ring_buffer_event *event;
2046 * Only the event that crossed the page boundary
2047 * must fill the old tail_page with padding.
2049 if (tail >= BUF_PAGE_SIZE) {
2051 * If the page was filled, then we still need
2052 * to update the real_end. Reset it to zero
2053 * and the reader will ignore it.
2055 if (tail == BUF_PAGE_SIZE)
2056 tail_page->real_end = 0;
2058 local_sub(length, &tail_page->write);
2062 event = __rb_page_index(tail_page, tail);
2063 kmemcheck_annotate_bitfield(event, bitfield);
2065 /* account for padding bytes */
2066 local_add(BUF_PAGE_SIZE - tail, &cpu_buffer->entries_bytes);
2069 * Save the original length to the meta data.
2070 * This will be used by the reader to add lost event
2073 tail_page->real_end = tail;
2076 * If this event is bigger than the minimum size, then
2077 * we need to be careful that we don't subtract the
2078 * write counter enough to allow another writer to slip
2080 * We put in a discarded commit instead, to make sure
2081 * that this space is not used again.
2083 * If we are less than the minimum size, we don't need to
2086 if (tail > (BUF_PAGE_SIZE - RB_EVNT_MIN_SIZE)) {
2087 /* No room for any events */
2089 /* Mark the rest of the page with padding */
2090 rb_event_set_padding(event);
2092 /* Set the write back to the previous setting */
2093 local_sub(length, &tail_page->write);
2097 /* Put in a discarded event */
2098 event->array[0] = (BUF_PAGE_SIZE - tail) - RB_EVNT_HDR_SIZE;
2099 event->type_len = RINGBUF_TYPE_PADDING;
2100 /* time delta must be non zero */
2101 event->time_delta = 1;
2103 /* Set write to end of buffer */
2104 length = (tail + length) - BUF_PAGE_SIZE;
2105 local_sub(length, &tail_page->write);
2109 * This is the slow path, force gcc not to inline it.
2111 static noinline struct ring_buffer_event *
2112 rb_move_tail(struct ring_buffer_per_cpu *cpu_buffer,
2113 unsigned long length, unsigned long tail,
2114 struct buffer_page *tail_page, u64 ts)
2116 struct buffer_page *commit_page = cpu_buffer->commit_page;
2117 struct ring_buffer *buffer = cpu_buffer->buffer;
2118 struct buffer_page *next_page;
2121 next_page = tail_page;
2123 rb_inc_page(cpu_buffer, &next_page);
2126 * If for some reason, we had an interrupt storm that made
2127 * it all the way around the buffer, bail, and warn
2130 if (unlikely(next_page == commit_page)) {
2131 local_inc(&cpu_buffer->commit_overrun);
2136 * This is where the fun begins!
2138 * We are fighting against races between a reader that
2139 * could be on another CPU trying to swap its reader
2140 * page with the buffer head.
2142 * We are also fighting against interrupts coming in and
2143 * moving the head or tail on us as well.
2145 * If the next page is the head page then we have filled
2146 * the buffer, unless the commit page is still on the
2149 if (rb_is_head_page(cpu_buffer, next_page, &tail_page->list)) {
2152 * If the commit is not on the reader page, then
2153 * move the header page.
2155 if (!rb_is_reader_page(cpu_buffer->commit_page)) {
2157 * If we are not in overwrite mode,
2158 * this is easy, just stop here.
2160 if (!(buffer->flags & RB_FL_OVERWRITE))
2163 ret = rb_handle_head_page(cpu_buffer,
2172 * We need to be careful here too. The
2173 * commit page could still be on the reader
2174 * page. We could have a small buffer, and
2175 * have filled up the buffer with events
2176 * from interrupts and such, and wrapped.
2178 * Note, if the tail page is also the on the
2179 * reader_page, we let it move out.
2181 if (unlikely((cpu_buffer->commit_page !=
2182 cpu_buffer->tail_page) &&
2183 (cpu_buffer->commit_page ==
2184 cpu_buffer->reader_page))) {
2185 local_inc(&cpu_buffer->commit_overrun);
2191 ret = rb_tail_page_update(cpu_buffer, tail_page, next_page);
2194 * Nested commits always have zero deltas, so
2195 * just reread the time stamp
2197 ts = rb_time_stamp(buffer);
2198 next_page->page->time_stamp = ts;
2203 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2205 /* fail and let the caller try again */
2206 return ERR_PTR(-EAGAIN);
2210 rb_reset_tail(cpu_buffer, tail_page, tail, length);
2215 static struct ring_buffer_event *
2216 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
2217 unsigned long length, u64 ts,
2218 u64 delta, int add_timestamp)
2220 struct buffer_page *tail_page;
2221 struct ring_buffer_event *event;
2222 unsigned long tail, write;
2225 * If the time delta since the last event is too big to
2226 * hold in the time field of the event, then we append a
2227 * TIME EXTEND event ahead of the data event.
2229 if (unlikely(add_timestamp))
2230 length += RB_LEN_TIME_EXTEND;
2232 tail_page = cpu_buffer->tail_page;
2233 write = local_add_return(length, &tail_page->write);
2235 /* set write to only the index of the write */
2236 write &= RB_WRITE_MASK;
2237 tail = write - length;
2239 /* See if we shot pass the end of this buffer page */
2240 if (unlikely(write > BUF_PAGE_SIZE))
2241 return rb_move_tail(cpu_buffer, length, tail,
2244 /* We reserved something on the buffer */
2246 event = __rb_page_index(tail_page, tail);
2247 kmemcheck_annotate_bitfield(event, bitfield);
2248 rb_update_event(cpu_buffer, event, length, add_timestamp, delta);
2250 local_inc(&tail_page->entries);
2253 * If this is the first commit on the page, then update
2257 tail_page->page->time_stamp = ts;
2259 /* account for these added bytes */
2260 local_add(length, &cpu_buffer->entries_bytes);
2266 rb_try_to_discard(struct ring_buffer_per_cpu *cpu_buffer,
2267 struct ring_buffer_event *event)
2269 unsigned long new_index, old_index;
2270 struct buffer_page *bpage;
2271 unsigned long index;
2274 new_index = rb_event_index(event);
2275 old_index = new_index + rb_event_ts_length(event);
2276 addr = (unsigned long)event;
2279 bpage = cpu_buffer->tail_page;
2281 if (bpage->page == (void *)addr && rb_page_write(bpage) == old_index) {
2282 unsigned long write_mask =
2283 local_read(&bpage->write) & ~RB_WRITE_MASK;
2284 unsigned long event_length = rb_event_length(event);
2286 * This is on the tail page. It is possible that
2287 * a write could come in and move the tail page
2288 * and write to the next page. That is fine
2289 * because we just shorten what is on this page.
2291 old_index += write_mask;
2292 new_index += write_mask;
2293 index = local_cmpxchg(&bpage->write, old_index, new_index);
2294 if (index == old_index) {
2295 /* update counters */
2296 local_sub(event_length, &cpu_buffer->entries_bytes);
2301 /* could not discard */
2305 static void rb_start_commit(struct ring_buffer_per_cpu *cpu_buffer)
2307 local_inc(&cpu_buffer->committing);
2308 local_inc(&cpu_buffer->commits);
2311 static inline void rb_end_commit(struct ring_buffer_per_cpu *cpu_buffer)
2313 unsigned long commits;
2315 if (RB_WARN_ON(cpu_buffer,
2316 !local_read(&cpu_buffer->committing)))
2320 commits = local_read(&cpu_buffer->commits);
2321 /* synchronize with interrupts */
2323 if (local_read(&cpu_buffer->committing) == 1)
2324 rb_set_commit_to_write(cpu_buffer);
2326 local_dec(&cpu_buffer->committing);
2328 /* synchronize with interrupts */
2332 * Need to account for interrupts coming in between the
2333 * updating of the commit page and the clearing of the
2334 * committing counter.
2336 if (unlikely(local_read(&cpu_buffer->commits) != commits) &&
2337 !local_read(&cpu_buffer->committing)) {
2338 local_inc(&cpu_buffer->committing);
2343 static struct ring_buffer_event *
2344 rb_reserve_next_event(struct ring_buffer *buffer,
2345 struct ring_buffer_per_cpu *cpu_buffer,
2346 unsigned long length)
2348 struct ring_buffer_event *event;
2354 rb_start_commit(cpu_buffer);
2356 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
2358 * Due to the ability to swap a cpu buffer from a buffer
2359 * it is possible it was swapped before we committed.
2360 * (committing stops a swap). We check for it here and
2361 * if it happened, we have to fail the write.
2364 if (unlikely(ACCESS_ONCE(cpu_buffer->buffer) != buffer)) {
2365 local_dec(&cpu_buffer->committing);
2366 local_dec(&cpu_buffer->commits);
2371 length = rb_calculate_event_length(length);
2377 * We allow for interrupts to reenter here and do a trace.
2378 * If one does, it will cause this original code to loop
2379 * back here. Even with heavy interrupts happening, this
2380 * should only happen a few times in a row. If this happens
2381 * 1000 times in a row, there must be either an interrupt
2382 * storm or we have something buggy.
2385 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
2388 ts = rb_time_stamp(cpu_buffer->buffer);
2389 diff = ts - cpu_buffer->write_stamp;
2391 /* make sure this diff is calculated here */
2394 /* Did the write stamp get updated already? */
2395 if (likely(ts >= cpu_buffer->write_stamp)) {
2397 if (unlikely(test_time_stamp(delta))) {
2398 int local_clock_stable = 1;
2399 #ifdef CONFIG_HAVE_UNSTABLE_SCHED_CLOCK
2400 local_clock_stable = sched_clock_stable;
2402 WARN_ONCE(delta > (1ULL << 59),
2403 KERN_WARNING "Delta way too big! %llu ts=%llu write stamp = %llu\n%s",
2404 (unsigned long long)delta,
2405 (unsigned long long)ts,
2406 (unsigned long long)cpu_buffer->write_stamp,
2407 local_clock_stable ? "" :
2408 "If you just came from a suspend/resume,\n"
2409 "please switch to the trace global clock:\n"
2410 " echo global > /sys/kernel/debug/tracing/trace_clock\n");
2415 event = __rb_reserve_next(cpu_buffer, length, ts,
2416 delta, add_timestamp);
2417 if (unlikely(PTR_ERR(event) == -EAGAIN))
2426 rb_end_commit(cpu_buffer);
2430 #ifdef CONFIG_TRACING
2432 #define TRACE_RECURSIVE_DEPTH 16
2434 /* Keep this code out of the fast path cache */
2435 static noinline void trace_recursive_fail(void)
2437 /* Disable all tracing before we do anything else */
2438 tracing_off_permanent();
2440 printk_once(KERN_WARNING "Tracing recursion: depth[%ld]:"
2441 "HC[%lu]:SC[%lu]:NMI[%lu]\n",
2442 trace_recursion_buffer(),
2443 hardirq_count() >> HARDIRQ_SHIFT,
2444 softirq_count() >> SOFTIRQ_SHIFT,
2450 static inline int trace_recursive_lock(void)
2452 trace_recursion_inc();
2454 if (likely(trace_recursion_buffer() < TRACE_RECURSIVE_DEPTH))
2457 trace_recursive_fail();
2462 static inline void trace_recursive_unlock(void)
2464 WARN_ON_ONCE(!trace_recursion_buffer());
2466 trace_recursion_dec();
2471 #define trace_recursive_lock() (0)
2472 #define trace_recursive_unlock() do { } while (0)
2477 * ring_buffer_lock_reserve - reserve a part of the buffer
2478 * @buffer: the ring buffer to reserve from
2479 * @length: the length of the data to reserve (excluding event header)
2481 * Returns a reseverd event on the ring buffer to copy directly to.
2482 * The user of this interface will need to get the body to write into
2483 * and can use the ring_buffer_event_data() interface.
2485 * The length is the length of the data needed, not the event length
2486 * which also includes the event header.
2488 * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
2489 * If NULL is returned, then nothing has been allocated or locked.
2491 struct ring_buffer_event *
2492 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
2494 struct ring_buffer_per_cpu *cpu_buffer;
2495 struct ring_buffer_event *event;
2498 if (ring_buffer_flags != RB_BUFFERS_ON)
2501 /* If we are tracing schedule, we don't want to recurse */
2502 preempt_disable_notrace();
2504 if (atomic_read(&buffer->record_disabled))
2507 if (trace_recursive_lock())
2510 cpu = raw_smp_processor_id();
2512 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2515 cpu_buffer = buffer->buffers[cpu];
2517 if (atomic_read(&cpu_buffer->record_disabled))
2520 if (length > BUF_MAX_DATA_SIZE)
2523 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2530 trace_recursive_unlock();
2533 preempt_enable_notrace();
2536 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
2539 rb_update_write_stamp(struct ring_buffer_per_cpu *cpu_buffer,
2540 struct ring_buffer_event *event)
2545 * The event first in the commit queue updates the
2548 if (rb_event_is_commit(cpu_buffer, event)) {
2550 * A commit event that is first on a page
2551 * updates the write timestamp with the page stamp
2553 if (!rb_event_index(event))
2554 cpu_buffer->write_stamp =
2555 cpu_buffer->commit_page->page->time_stamp;
2556 else if (event->type_len == RINGBUF_TYPE_TIME_EXTEND) {
2557 delta = event->array[0];
2559 delta += event->time_delta;
2560 cpu_buffer->write_stamp += delta;
2562 cpu_buffer->write_stamp += event->time_delta;
2566 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
2567 struct ring_buffer_event *event)
2569 local_inc(&cpu_buffer->entries);
2570 rb_update_write_stamp(cpu_buffer, event);
2571 rb_end_commit(cpu_buffer);
2575 * ring_buffer_unlock_commit - commit a reserved
2576 * @buffer: The buffer to commit to
2577 * @event: The event pointer to commit.
2579 * This commits the data to the ring buffer, and releases any locks held.
2581 * Must be paired with ring_buffer_lock_reserve.
2583 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
2584 struct ring_buffer_event *event)
2586 struct ring_buffer_per_cpu *cpu_buffer;
2587 int cpu = raw_smp_processor_id();
2589 cpu_buffer = buffer->buffers[cpu];
2591 rb_commit(cpu_buffer, event);
2593 trace_recursive_unlock();
2595 preempt_enable_notrace();
2599 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
2601 static inline void rb_event_discard(struct ring_buffer_event *event)
2603 if (event->type_len == RINGBUF_TYPE_TIME_EXTEND)
2604 event = skip_time_extend(event);
2606 /* array[0] holds the actual length for the discarded event */
2607 event->array[0] = rb_event_data_length(event) - RB_EVNT_HDR_SIZE;
2608 event->type_len = RINGBUF_TYPE_PADDING;
2609 /* time delta must be non zero */
2610 if (!event->time_delta)
2611 event->time_delta = 1;
2615 * Decrement the entries to the page that an event is on.
2616 * The event does not even need to exist, only the pointer
2617 * to the page it is on. This may only be called before the commit
2621 rb_decrement_entry(struct ring_buffer_per_cpu *cpu_buffer,
2622 struct ring_buffer_event *event)
2624 unsigned long addr = (unsigned long)event;
2625 struct buffer_page *bpage = cpu_buffer->commit_page;
2626 struct buffer_page *start;
2630 /* Do the likely case first */
2631 if (likely(bpage->page == (void *)addr)) {
2632 local_dec(&bpage->entries);
2637 * Because the commit page may be on the reader page we
2638 * start with the next page and check the end loop there.
2640 rb_inc_page(cpu_buffer, &bpage);
2643 if (bpage->page == (void *)addr) {
2644 local_dec(&bpage->entries);
2647 rb_inc_page(cpu_buffer, &bpage);
2648 } while (bpage != start);
2650 /* commit not part of this buffer?? */
2651 RB_WARN_ON(cpu_buffer, 1);
2655 * ring_buffer_commit_discard - discard an event that has not been committed
2656 * @buffer: the ring buffer
2657 * @event: non committed event to discard
2659 * Sometimes an event that is in the ring buffer needs to be ignored.
2660 * This function lets the user discard an event in the ring buffer
2661 * and then that event will not be read later.
2663 * This function only works if it is called before the the item has been
2664 * committed. It will try to free the event from the ring buffer
2665 * if another event has not been added behind it.
2667 * If another event has been added behind it, it will set the event
2668 * up as discarded, and perform the commit.
2670 * If this function is called, do not call ring_buffer_unlock_commit on
2673 void ring_buffer_discard_commit(struct ring_buffer *buffer,
2674 struct ring_buffer_event *event)
2676 struct ring_buffer_per_cpu *cpu_buffer;
2679 /* The event is discarded regardless */
2680 rb_event_discard(event);
2682 cpu = smp_processor_id();
2683 cpu_buffer = buffer->buffers[cpu];
2686 * This must only be called if the event has not been
2687 * committed yet. Thus we can assume that preemption
2688 * is still disabled.
2690 RB_WARN_ON(buffer, !local_read(&cpu_buffer->committing));
2692 rb_decrement_entry(cpu_buffer, event);
2693 if (rb_try_to_discard(cpu_buffer, event))
2697 * The commit is still visible by the reader, so we
2698 * must still update the timestamp.
2700 rb_update_write_stamp(cpu_buffer, event);
2702 rb_end_commit(cpu_buffer);
2704 trace_recursive_unlock();
2706 preempt_enable_notrace();
2709 EXPORT_SYMBOL_GPL(ring_buffer_discard_commit);
2712 * ring_buffer_write - write data to the buffer without reserving
2713 * @buffer: The ring buffer to write to.
2714 * @length: The length of the data being written (excluding the event header)
2715 * @data: The data to write to the buffer.
2717 * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
2718 * one function. If you already have the data to write to the buffer, it
2719 * may be easier to simply call this function.
2721 * Note, like ring_buffer_lock_reserve, the length is the length of the data
2722 * and not the length of the event which would hold the header.
2724 int ring_buffer_write(struct ring_buffer *buffer,
2725 unsigned long length,
2728 struct ring_buffer_per_cpu *cpu_buffer;
2729 struct ring_buffer_event *event;
2734 if (ring_buffer_flags != RB_BUFFERS_ON)
2737 preempt_disable_notrace();
2739 if (atomic_read(&buffer->record_disabled))
2742 cpu = raw_smp_processor_id();
2744 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2747 cpu_buffer = buffer->buffers[cpu];
2749 if (atomic_read(&cpu_buffer->record_disabled))
2752 if (length > BUF_MAX_DATA_SIZE)
2755 event = rb_reserve_next_event(buffer, cpu_buffer, length);
2759 body = rb_event_data(event);
2761 memcpy(body, data, length);
2763 rb_commit(cpu_buffer, event);
2767 preempt_enable_notrace();
2771 EXPORT_SYMBOL_GPL(ring_buffer_write);
2773 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
2775 struct buffer_page *reader = cpu_buffer->reader_page;
2776 struct buffer_page *head = rb_set_head_page(cpu_buffer);
2777 struct buffer_page *commit = cpu_buffer->commit_page;
2779 /* In case of error, head will be NULL */
2780 if (unlikely(!head))
2783 return reader->read == rb_page_commit(reader) &&
2784 (commit == reader ||
2786 head->read == rb_page_commit(commit)));
2790 * ring_buffer_record_disable - stop all writes into the buffer
2791 * @buffer: The ring buffer to stop writes to.
2793 * This prevents all writes to the buffer. Any attempt to write
2794 * to the buffer after this will fail and return NULL.
2796 * The caller should call synchronize_sched() after this.
2798 void ring_buffer_record_disable(struct ring_buffer *buffer)
2800 atomic_inc(&buffer->record_disabled);
2802 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
2805 * ring_buffer_record_enable - enable writes to the buffer
2806 * @buffer: The ring buffer to enable writes
2808 * Note, multiple disables will need the same number of enables
2809 * to truly enable the writing (much like preempt_disable).
2811 void ring_buffer_record_enable(struct ring_buffer *buffer)
2813 atomic_dec(&buffer->record_disabled);
2815 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
2818 * ring_buffer_record_off - stop all writes into the buffer
2819 * @buffer: The ring buffer to stop writes to.
2821 * This prevents all writes to the buffer. Any attempt to write
2822 * to the buffer after this will fail and return NULL.
2824 * This is different than ring_buffer_record_disable() as
2825 * it works like an on/off switch, where as the disable() version
2826 * must be paired with a enable().
2828 void ring_buffer_record_off(struct ring_buffer *buffer)
2831 unsigned int new_rd;
2834 rd = atomic_read(&buffer->record_disabled);
2835 new_rd = rd | RB_BUFFER_OFF;
2836 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
2838 EXPORT_SYMBOL_GPL(ring_buffer_record_off);
2841 * ring_buffer_record_on - restart writes into the buffer
2842 * @buffer: The ring buffer to start writes to.
2844 * This enables all writes to the buffer that was disabled by
2845 * ring_buffer_record_off().
2847 * This is different than ring_buffer_record_enable() as
2848 * it works like an on/off switch, where as the enable() version
2849 * must be paired with a disable().
2851 void ring_buffer_record_on(struct ring_buffer *buffer)
2854 unsigned int new_rd;
2857 rd = atomic_read(&buffer->record_disabled);
2858 new_rd = rd & ~RB_BUFFER_OFF;
2859 } while (atomic_cmpxchg(&buffer->record_disabled, rd, new_rd) != rd);
2861 EXPORT_SYMBOL_GPL(ring_buffer_record_on);
2864 * ring_buffer_record_is_on - return true if the ring buffer can write
2865 * @buffer: The ring buffer to see if write is enabled
2867 * Returns true if the ring buffer is in a state that it accepts writes.
2869 int ring_buffer_record_is_on(struct ring_buffer *buffer)
2871 return !atomic_read(&buffer->record_disabled);
2875 * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
2876 * @buffer: The ring buffer to stop writes to.
2877 * @cpu: The CPU buffer to stop
2879 * This prevents all writes to the buffer. Any attempt to write
2880 * to the buffer after this will fail and return NULL.
2882 * The caller should call synchronize_sched() after this.
2884 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
2886 struct ring_buffer_per_cpu *cpu_buffer;
2888 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2891 cpu_buffer = buffer->buffers[cpu];
2892 atomic_inc(&cpu_buffer->record_disabled);
2894 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
2897 * ring_buffer_record_enable_cpu - enable writes to the buffer
2898 * @buffer: The ring buffer to enable writes
2899 * @cpu: The CPU to enable.
2901 * Note, multiple disables will need the same number of enables
2902 * to truly enable the writing (much like preempt_disable).
2904 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
2906 struct ring_buffer_per_cpu *cpu_buffer;
2908 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2911 cpu_buffer = buffer->buffers[cpu];
2912 atomic_dec(&cpu_buffer->record_disabled);
2914 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
2917 * The total entries in the ring buffer is the running counter
2918 * of entries entered into the ring buffer, minus the sum of
2919 * the entries read from the ring buffer and the number of
2920 * entries that were overwritten.
2922 static inline unsigned long
2923 rb_num_of_entries(struct ring_buffer_per_cpu *cpu_buffer)
2925 return local_read(&cpu_buffer->entries) -
2926 (local_read(&cpu_buffer->overrun) + cpu_buffer->read);
2930 * ring_buffer_oldest_event_ts - get the oldest event timestamp from the buffer
2931 * @buffer: The ring buffer
2932 * @cpu: The per CPU buffer to read from.
2934 unsigned long ring_buffer_oldest_event_ts(struct ring_buffer *buffer, int cpu)
2936 unsigned long flags;
2937 struct ring_buffer_per_cpu *cpu_buffer;
2938 struct buffer_page *bpage;
2939 unsigned long ret = 0;
2941 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2944 cpu_buffer = buffer->buffers[cpu];
2945 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2947 * if the tail is on reader_page, oldest time stamp is on the reader
2950 if (cpu_buffer->tail_page == cpu_buffer->reader_page)
2951 bpage = cpu_buffer->reader_page;
2953 bpage = rb_set_head_page(cpu_buffer);
2955 ret = bpage->page->time_stamp;
2956 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2960 EXPORT_SYMBOL_GPL(ring_buffer_oldest_event_ts);
2963 * ring_buffer_bytes_cpu - get the number of bytes consumed in a cpu buffer
2964 * @buffer: The ring buffer
2965 * @cpu: The per CPU buffer to read from.
2967 unsigned long ring_buffer_bytes_cpu(struct ring_buffer *buffer, int cpu)
2969 struct ring_buffer_per_cpu *cpu_buffer;
2972 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2975 cpu_buffer = buffer->buffers[cpu];
2976 ret = local_read(&cpu_buffer->entries_bytes) - cpu_buffer->read_bytes;
2980 EXPORT_SYMBOL_GPL(ring_buffer_bytes_cpu);
2983 * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
2984 * @buffer: The ring buffer
2985 * @cpu: The per CPU buffer to get the entries from.
2987 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
2989 struct ring_buffer_per_cpu *cpu_buffer;
2991 if (!cpumask_test_cpu(cpu, buffer->cpumask))
2994 cpu_buffer = buffer->buffers[cpu];
2996 return rb_num_of_entries(cpu_buffer);
2998 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
3001 * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
3002 * @buffer: The ring buffer
3003 * @cpu: The per CPU buffer to get the number of overruns from
3005 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
3007 struct ring_buffer_per_cpu *cpu_buffer;
3010 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3013 cpu_buffer = buffer->buffers[cpu];
3014 ret = local_read(&cpu_buffer->overrun);
3018 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
3021 * ring_buffer_commit_overrun_cpu - get the number of overruns caused by commits
3022 * @buffer: The ring buffer
3023 * @cpu: The per CPU buffer to get the number of overruns from
3026 ring_buffer_commit_overrun_cpu(struct ring_buffer *buffer, int cpu)
3028 struct ring_buffer_per_cpu *cpu_buffer;
3031 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3034 cpu_buffer = buffer->buffers[cpu];
3035 ret = local_read(&cpu_buffer->commit_overrun);
3039 EXPORT_SYMBOL_GPL(ring_buffer_commit_overrun_cpu);
3042 * ring_buffer_entries - get the number of entries in a buffer
3043 * @buffer: The ring buffer
3045 * Returns the total number of entries in the ring buffer
3048 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
3050 struct ring_buffer_per_cpu *cpu_buffer;
3051 unsigned long entries = 0;
3054 /* if you care about this being correct, lock the buffer */
3055 for_each_buffer_cpu(buffer, cpu) {
3056 cpu_buffer = buffer->buffers[cpu];
3057 entries += rb_num_of_entries(cpu_buffer);
3062 EXPORT_SYMBOL_GPL(ring_buffer_entries);
3065 * ring_buffer_overruns - get the number of overruns in buffer
3066 * @buffer: The ring buffer
3068 * Returns the total number of overruns in the ring buffer
3071 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
3073 struct ring_buffer_per_cpu *cpu_buffer;
3074 unsigned long overruns = 0;
3077 /* if you care about this being correct, lock the buffer */
3078 for_each_buffer_cpu(buffer, cpu) {
3079 cpu_buffer = buffer->buffers[cpu];
3080 overruns += local_read(&cpu_buffer->overrun);
3085 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
3087 static void rb_iter_reset(struct ring_buffer_iter *iter)
3089 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3091 /* Iterator usage is expected to have record disabled */
3092 if (list_empty(&cpu_buffer->reader_page->list)) {
3093 iter->head_page = rb_set_head_page(cpu_buffer);
3094 if (unlikely(!iter->head_page))
3096 iter->head = iter->head_page->read;
3098 iter->head_page = cpu_buffer->reader_page;
3099 iter->head = cpu_buffer->reader_page->read;
3102 iter->read_stamp = cpu_buffer->read_stamp;
3104 iter->read_stamp = iter->head_page->page->time_stamp;
3105 iter->cache_reader_page = cpu_buffer->reader_page;
3106 iter->cache_read = cpu_buffer->read;
3110 * ring_buffer_iter_reset - reset an iterator
3111 * @iter: The iterator to reset
3113 * Resets the iterator, so that it will start from the beginning
3116 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
3118 struct ring_buffer_per_cpu *cpu_buffer;
3119 unsigned long flags;
3124 cpu_buffer = iter->cpu_buffer;
3126 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3127 rb_iter_reset(iter);
3128 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3130 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
3133 * ring_buffer_iter_empty - check if an iterator has no more to read
3134 * @iter: The iterator to check
3136 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
3138 struct ring_buffer_per_cpu *cpu_buffer;
3140 cpu_buffer = iter->cpu_buffer;
3142 return iter->head_page == cpu_buffer->commit_page &&
3143 iter->head == rb_commit_index(cpu_buffer);
3145 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
3148 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
3149 struct ring_buffer_event *event)
3153 switch (event->type_len) {
3154 case RINGBUF_TYPE_PADDING:
3157 case RINGBUF_TYPE_TIME_EXTEND:
3158 delta = event->array[0];
3160 delta += event->time_delta;
3161 cpu_buffer->read_stamp += delta;
3164 case RINGBUF_TYPE_TIME_STAMP:
3165 /* FIXME: not implemented */
3168 case RINGBUF_TYPE_DATA:
3169 cpu_buffer->read_stamp += event->time_delta;
3179 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
3180 struct ring_buffer_event *event)
3184 switch (event->type_len) {
3185 case RINGBUF_TYPE_PADDING:
3188 case RINGBUF_TYPE_TIME_EXTEND:
3189 delta = event->array[0];
3191 delta += event->time_delta;
3192 iter->read_stamp += delta;
3195 case RINGBUF_TYPE_TIME_STAMP:
3196 /* FIXME: not implemented */
3199 case RINGBUF_TYPE_DATA:
3200 iter->read_stamp += event->time_delta;
3209 static struct buffer_page *
3210 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
3212 struct buffer_page *reader = NULL;
3213 unsigned long overwrite;
3214 unsigned long flags;
3218 local_irq_save(flags);
3219 arch_spin_lock(&cpu_buffer->lock);
3223 * This should normally only loop twice. But because the
3224 * start of the reader inserts an empty page, it causes
3225 * a case where we will loop three times. There should be no
3226 * reason to loop four times (that I know of).
3228 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
3233 reader = cpu_buffer->reader_page;
3235 /* If there's more to read, return this page */
3236 if (cpu_buffer->reader_page->read < rb_page_size(reader))
3239 /* Never should we have an index greater than the size */
3240 if (RB_WARN_ON(cpu_buffer,
3241 cpu_buffer->reader_page->read > rb_page_size(reader)))
3244 /* check if we caught up to the tail */
3246 if (cpu_buffer->commit_page == cpu_buffer->reader_page)
3249 /* Don't bother swapping if the ring buffer is empty */
3250 if (rb_num_of_entries(cpu_buffer) == 0)
3254 * Reset the reader page to size zero.
3256 local_set(&cpu_buffer->reader_page->write, 0);
3257 local_set(&cpu_buffer->reader_page->entries, 0);
3258 local_set(&cpu_buffer->reader_page->page->commit, 0);
3259 cpu_buffer->reader_page->real_end = 0;
3263 * Splice the empty reader page into the list around the head.
3265 reader = rb_set_head_page(cpu_buffer);
3268 cpu_buffer->reader_page->list.next = rb_list_head(reader->list.next);
3269 cpu_buffer->reader_page->list.prev = reader->list.prev;
3272 * cpu_buffer->pages just needs to point to the buffer, it
3273 * has no specific buffer page to point to. Lets move it out
3274 * of our way so we don't accidentally swap it.
3276 cpu_buffer->pages = reader->list.prev;
3278 /* The reader page will be pointing to the new head */
3279 rb_set_list_to_head(cpu_buffer, &cpu_buffer->reader_page->list);
3282 * We want to make sure we read the overruns after we set up our
3283 * pointers to the next object. The writer side does a
3284 * cmpxchg to cross pages which acts as the mb on the writer
3285 * side. Note, the reader will constantly fail the swap
3286 * while the writer is updating the pointers, so this
3287 * guarantees that the overwrite recorded here is the one we
3288 * want to compare with the last_overrun.
3291 overwrite = local_read(&(cpu_buffer->overrun));
3294 * Here's the tricky part.
3296 * We need to move the pointer past the header page.
3297 * But we can only do that if a writer is not currently
3298 * moving it. The page before the header page has the
3299 * flag bit '1' set if it is pointing to the page we want.
3300 * but if the writer is in the process of moving it
3301 * than it will be '2' or already moved '0'.
3304 ret = rb_head_page_replace(reader, cpu_buffer->reader_page);
3307 * If we did not convert it, then we must try again.
3313 * Yeah! We succeeded in replacing the page.
3315 * Now make the new head point back to the reader page.
3317 rb_list_head(reader->list.next)->prev = &cpu_buffer->reader_page->list;
3318 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
3320 /* Finally update the reader page to the new head */
3321 cpu_buffer->reader_page = reader;
3322 rb_reset_reader_page(cpu_buffer);
3324 if (overwrite != cpu_buffer->last_overrun) {
3325 cpu_buffer->lost_events = overwrite - cpu_buffer->last_overrun;
3326 cpu_buffer->last_overrun = overwrite;
3332 arch_spin_unlock(&cpu_buffer->lock);
3333 local_irq_restore(flags);
3338 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
3340 struct ring_buffer_event *event;
3341 struct buffer_page *reader;
3344 reader = rb_get_reader_page(cpu_buffer);
3346 /* This function should not be called when buffer is empty */
3347 if (RB_WARN_ON(cpu_buffer, !reader))
3350 event = rb_reader_event(cpu_buffer);
3352 if (event->type_len <= RINGBUF_TYPE_DATA_TYPE_LEN_MAX)
3355 rb_update_read_stamp(cpu_buffer, event);
3357 length = rb_event_length(event);
3358 cpu_buffer->reader_page->read += length;
3361 static void rb_advance_iter(struct ring_buffer_iter *iter)
3363 struct ring_buffer_per_cpu *cpu_buffer;
3364 struct ring_buffer_event *event;
3367 cpu_buffer = iter->cpu_buffer;
3370 * Check if we are at the end of the buffer.
3372 if (iter->head >= rb_page_size(iter->head_page)) {
3373 /* discarded commits can make the page empty */
3374 if (iter->head_page == cpu_buffer->commit_page)
3380 event = rb_iter_head_event(iter);
3382 length = rb_event_length(event);
3385 * This should not be called to advance the header if we are
3386 * at the tail of the buffer.
3388 if (RB_WARN_ON(cpu_buffer,
3389 (iter->head_page == cpu_buffer->commit_page) &&
3390 (iter->head + length > rb_commit_index(cpu_buffer))))
3393 rb_update_iter_read_stamp(iter, event);
3395 iter->head += length;
3397 /* check for end of page padding */
3398 if ((iter->head >= rb_page_size(iter->head_page)) &&
3399 (iter->head_page != cpu_buffer->commit_page))
3400 rb_advance_iter(iter);
3403 static int rb_lost_events(struct ring_buffer_per_cpu *cpu_buffer)
3405 return cpu_buffer->lost_events;
3408 static struct ring_buffer_event *
3409 rb_buffer_peek(struct ring_buffer_per_cpu *cpu_buffer, u64 *ts,
3410 unsigned long *lost_events)
3412 struct ring_buffer_event *event;
3413 struct buffer_page *reader;
3418 * We repeat when a time extend is encountered.
3419 * Since the time extend is always attached to a data event,
3420 * we should never loop more than once.
3421 * (We never hit the following condition more than twice).
3423 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3426 reader = rb_get_reader_page(cpu_buffer);
3430 event = rb_reader_event(cpu_buffer);
3432 switch (event->type_len) {
3433 case RINGBUF_TYPE_PADDING:
3434 if (rb_null_event(event))
3435 RB_WARN_ON(cpu_buffer, 1);
3437 * Because the writer could be discarding every
3438 * event it creates (which would probably be bad)
3439 * if we were to go back to "again" then we may never
3440 * catch up, and will trigger the warn on, or lock
3441 * the box. Return the padding, and we will release
3442 * the current locks, and try again.
3446 case RINGBUF_TYPE_TIME_EXTEND:
3447 /* Internal data, OK to advance */
3448 rb_advance_reader(cpu_buffer);
3451 case RINGBUF_TYPE_TIME_STAMP:
3452 /* FIXME: not implemented */
3453 rb_advance_reader(cpu_buffer);
3456 case RINGBUF_TYPE_DATA:
3458 *ts = cpu_buffer->read_stamp + event->time_delta;
3459 ring_buffer_normalize_time_stamp(cpu_buffer->buffer,
3460 cpu_buffer->cpu, ts);
3463 *lost_events = rb_lost_events(cpu_buffer);
3472 EXPORT_SYMBOL_GPL(ring_buffer_peek);
3474 static struct ring_buffer_event *
3475 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3477 struct ring_buffer *buffer;
3478 struct ring_buffer_per_cpu *cpu_buffer;
3479 struct ring_buffer_event *event;
3482 cpu_buffer = iter->cpu_buffer;
3483 buffer = cpu_buffer->buffer;
3486 * Check if someone performed a consuming read to
3487 * the buffer. A consuming read invalidates the iterator
3488 * and we need to reset the iterator in this case.
3490 if (unlikely(iter->cache_read != cpu_buffer->read ||
3491 iter->cache_reader_page != cpu_buffer->reader_page))
3492 rb_iter_reset(iter);
3495 if (ring_buffer_iter_empty(iter))
3499 * We repeat when a time extend is encountered.
3500 * Since the time extend is always attached to a data event,
3501 * we should never loop more than once.
3502 * (We never hit the following condition more than twice).
3504 if (RB_WARN_ON(cpu_buffer, ++nr_loops > 2))
3507 if (rb_per_cpu_empty(cpu_buffer))
3510 if (iter->head >= local_read(&iter->head_page->page->commit)) {
3515 event = rb_iter_head_event(iter);
3517 switch (event->type_len) {
3518 case RINGBUF_TYPE_PADDING:
3519 if (rb_null_event(event)) {
3523 rb_advance_iter(iter);
3526 case RINGBUF_TYPE_TIME_EXTEND:
3527 /* Internal data, OK to advance */
3528 rb_advance_iter(iter);
3531 case RINGBUF_TYPE_TIME_STAMP:
3532 /* FIXME: not implemented */
3533 rb_advance_iter(iter);
3536 case RINGBUF_TYPE_DATA:
3538 *ts = iter->read_stamp + event->time_delta;
3539 ring_buffer_normalize_time_stamp(buffer,
3540 cpu_buffer->cpu, ts);
3550 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
3552 static inline int rb_ok_to_lock(void)
3555 * If an NMI die dumps out the content of the ring buffer
3556 * do not grab locks. We also permanently disable the ring
3557 * buffer too. A one time deal is all you get from reading
3558 * the ring buffer from an NMI.
3560 if (likely(!in_nmi()))
3563 tracing_off_permanent();
3568 * ring_buffer_peek - peek at the next event to be read
3569 * @buffer: The ring buffer to read
3570 * @cpu: The cpu to peak at
3571 * @ts: The timestamp counter of this event.
3572 * @lost_events: a variable to store if events were lost (may be NULL)
3574 * This will return the event that will be read next, but does
3575 * not consume the data.
3577 struct ring_buffer_event *
3578 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts,
3579 unsigned long *lost_events)
3581 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3582 struct ring_buffer_event *event;
3583 unsigned long flags;
3586 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3589 dolock = rb_ok_to_lock();
3591 local_irq_save(flags);
3593 raw_spin_lock(&cpu_buffer->reader_lock);
3594 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3595 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3596 rb_advance_reader(cpu_buffer);
3598 raw_spin_unlock(&cpu_buffer->reader_lock);
3599 local_irq_restore(flags);
3601 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3608 * ring_buffer_iter_peek - peek at the next event to be read
3609 * @iter: The ring buffer iterator
3610 * @ts: The timestamp counter of this event.
3612 * This will return the event that will be read next, but does
3613 * not increment the iterator.
3615 struct ring_buffer_event *
3616 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
3618 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3619 struct ring_buffer_event *event;
3620 unsigned long flags;
3623 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3624 event = rb_iter_peek(iter, ts);
3625 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3627 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3634 * ring_buffer_consume - return an event and consume it
3635 * @buffer: The ring buffer to get the next event from
3636 * @cpu: the cpu to read the buffer from
3637 * @ts: a variable to store the timestamp (may be NULL)
3638 * @lost_events: a variable to store if events were lost (may be NULL)
3640 * Returns the next event in the ring buffer, and that event is consumed.
3641 * Meaning, that sequential reads will keep returning a different event,
3642 * and eventually empty the ring buffer if the producer is slower.
3644 struct ring_buffer_event *
3645 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts,
3646 unsigned long *lost_events)
3648 struct ring_buffer_per_cpu *cpu_buffer;
3649 struct ring_buffer_event *event = NULL;
3650 unsigned long flags;
3653 dolock = rb_ok_to_lock();
3656 /* might be called in atomic */
3659 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3662 cpu_buffer = buffer->buffers[cpu];
3663 local_irq_save(flags);
3665 raw_spin_lock(&cpu_buffer->reader_lock);
3667 event = rb_buffer_peek(cpu_buffer, ts, lost_events);
3669 cpu_buffer->lost_events = 0;
3670 rb_advance_reader(cpu_buffer);
3674 raw_spin_unlock(&cpu_buffer->reader_lock);
3675 local_irq_restore(flags);
3680 if (event && event->type_len == RINGBUF_TYPE_PADDING)
3685 EXPORT_SYMBOL_GPL(ring_buffer_consume);
3688 * ring_buffer_read_prepare - Prepare for a non consuming read of the buffer
3689 * @buffer: The ring buffer to read from
3690 * @cpu: The cpu buffer to iterate over
3692 * This performs the initial preparations necessary to iterate
3693 * through the buffer. Memory is allocated, buffer recording
3694 * is disabled, and the iterator pointer is returned to the caller.
3696 * Disabling buffer recordng prevents the reading from being
3697 * corrupted. This is not a consuming read, so a producer is not
3700 * After a sequence of ring_buffer_read_prepare calls, the user is
3701 * expected to make at least one call to ring_buffer_prepare_sync.
3702 * Afterwards, ring_buffer_read_start is invoked to get things going
3705 * This overall must be paired with ring_buffer_finish.
3707 struct ring_buffer_iter *
3708 ring_buffer_read_prepare(struct ring_buffer *buffer, int cpu)
3710 struct ring_buffer_per_cpu *cpu_buffer;
3711 struct ring_buffer_iter *iter;
3713 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3716 iter = kmalloc(sizeof(*iter), GFP_KERNEL);
3720 cpu_buffer = buffer->buffers[cpu];
3722 iter->cpu_buffer = cpu_buffer;
3724 atomic_inc(&buffer->resize_disabled);
3725 atomic_inc(&cpu_buffer->record_disabled);
3729 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare);
3732 * ring_buffer_read_prepare_sync - Synchronize a set of prepare calls
3734 * All previously invoked ring_buffer_read_prepare calls to prepare
3735 * iterators will be synchronized. Afterwards, read_buffer_read_start
3736 * calls on those iterators are allowed.
3739 ring_buffer_read_prepare_sync(void)
3741 synchronize_sched();
3743 EXPORT_SYMBOL_GPL(ring_buffer_read_prepare_sync);
3746 * ring_buffer_read_start - start a non consuming read of the buffer
3747 * @iter: The iterator returned by ring_buffer_read_prepare
3749 * This finalizes the startup of an iteration through the buffer.
3750 * The iterator comes from a call to ring_buffer_read_prepare and
3751 * an intervening ring_buffer_read_prepare_sync must have been
3754 * Must be paired with ring_buffer_finish.
3757 ring_buffer_read_start(struct ring_buffer_iter *iter)
3759 struct ring_buffer_per_cpu *cpu_buffer;
3760 unsigned long flags;
3765 cpu_buffer = iter->cpu_buffer;
3767 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3768 arch_spin_lock(&cpu_buffer->lock);
3769 rb_iter_reset(iter);
3770 arch_spin_unlock(&cpu_buffer->lock);
3771 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3773 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
3776 * ring_buffer_finish - finish reading the iterator of the buffer
3777 * @iter: The iterator retrieved by ring_buffer_start
3779 * This re-enables the recording to the buffer, and frees the
3783 ring_buffer_read_finish(struct ring_buffer_iter *iter)
3785 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3788 * Ring buffer is disabled from recording, here's a good place
3789 * to check the integrity of the ring buffer.
3791 rb_check_pages(cpu_buffer);
3793 atomic_dec(&cpu_buffer->record_disabled);
3794 atomic_dec(&cpu_buffer->buffer->resize_disabled);
3797 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
3800 * ring_buffer_read - read the next item in the ring buffer by the iterator
3801 * @iter: The ring buffer iterator
3802 * @ts: The time stamp of the event read.
3804 * This reads the next event in the ring buffer and increments the iterator.
3806 struct ring_buffer_event *
3807 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
3809 struct ring_buffer_event *event;
3810 struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
3811 unsigned long flags;
3813 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3815 event = rb_iter_peek(iter, ts);
3819 if (event->type_len == RINGBUF_TYPE_PADDING)
3822 rb_advance_iter(iter);
3824 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3828 EXPORT_SYMBOL_GPL(ring_buffer_read);
3831 * ring_buffer_size - return the size of the ring buffer (in bytes)
3832 * @buffer: The ring buffer.
3834 unsigned long ring_buffer_size(struct ring_buffer *buffer, int cpu)
3837 * Earlier, this method returned
3838 * BUF_PAGE_SIZE * buffer->nr_pages
3839 * Since the nr_pages field is now removed, we have converted this to
3840 * return the per cpu buffer value.
3842 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3845 return BUF_PAGE_SIZE * buffer->buffers[cpu]->nr_pages;
3847 EXPORT_SYMBOL_GPL(ring_buffer_size);
3850 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
3852 rb_head_page_deactivate(cpu_buffer);
3854 cpu_buffer->head_page
3855 = list_entry(cpu_buffer->pages, struct buffer_page, list);
3856 local_set(&cpu_buffer->head_page->write, 0);
3857 local_set(&cpu_buffer->head_page->entries, 0);
3858 local_set(&cpu_buffer->head_page->page->commit, 0);
3860 cpu_buffer->head_page->read = 0;
3862 cpu_buffer->tail_page = cpu_buffer->head_page;
3863 cpu_buffer->commit_page = cpu_buffer->head_page;
3865 INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
3866 INIT_LIST_HEAD(&cpu_buffer->new_pages);
3867 local_set(&cpu_buffer->reader_page->write, 0);
3868 local_set(&cpu_buffer->reader_page->entries, 0);
3869 local_set(&cpu_buffer->reader_page->page->commit, 0);
3870 cpu_buffer->reader_page->read = 0;
3872 local_set(&cpu_buffer->commit_overrun, 0);
3873 local_set(&cpu_buffer->entries_bytes, 0);
3874 local_set(&cpu_buffer->overrun, 0);
3875 local_set(&cpu_buffer->entries, 0);
3876 local_set(&cpu_buffer->committing, 0);
3877 local_set(&cpu_buffer->commits, 0);
3878 cpu_buffer->read = 0;
3879 cpu_buffer->read_bytes = 0;
3881 cpu_buffer->write_stamp = 0;
3882 cpu_buffer->read_stamp = 0;
3884 cpu_buffer->lost_events = 0;
3885 cpu_buffer->last_overrun = 0;
3887 rb_head_page_activate(cpu_buffer);
3891 * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
3892 * @buffer: The ring buffer to reset a per cpu buffer of
3893 * @cpu: The CPU buffer to be reset
3895 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
3897 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
3898 unsigned long flags;
3900 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3903 atomic_inc(&buffer->resize_disabled);
3904 atomic_inc(&cpu_buffer->record_disabled);
3906 /* Make sure all commits have finished */
3907 synchronize_sched();
3909 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
3911 if (RB_WARN_ON(cpu_buffer, local_read(&cpu_buffer->committing)))
3914 arch_spin_lock(&cpu_buffer->lock);
3916 rb_reset_cpu(cpu_buffer);
3918 arch_spin_unlock(&cpu_buffer->lock);
3921 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
3923 atomic_dec(&cpu_buffer->record_disabled);
3924 atomic_dec(&buffer->resize_disabled);
3926 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
3929 * ring_buffer_reset - reset a ring buffer
3930 * @buffer: The ring buffer to reset all cpu buffers
3932 void ring_buffer_reset(struct ring_buffer *buffer)
3936 for_each_buffer_cpu(buffer, cpu)
3937 ring_buffer_reset_cpu(buffer, cpu);
3939 EXPORT_SYMBOL_GPL(ring_buffer_reset);
3942 * rind_buffer_empty - is the ring buffer empty?
3943 * @buffer: The ring buffer to test
3945 int ring_buffer_empty(struct ring_buffer *buffer)
3947 struct ring_buffer_per_cpu *cpu_buffer;
3948 unsigned long flags;
3953 dolock = rb_ok_to_lock();
3955 /* yes this is racy, but if you don't like the race, lock the buffer */
3956 for_each_buffer_cpu(buffer, cpu) {
3957 cpu_buffer = buffer->buffers[cpu];
3958 local_irq_save(flags);
3960 raw_spin_lock(&cpu_buffer->reader_lock);
3961 ret = rb_per_cpu_empty(cpu_buffer);
3963 raw_spin_unlock(&cpu_buffer->reader_lock);
3964 local_irq_restore(flags);
3972 EXPORT_SYMBOL_GPL(ring_buffer_empty);
3975 * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
3976 * @buffer: The ring buffer
3977 * @cpu: The CPU buffer to test
3979 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
3981 struct ring_buffer_per_cpu *cpu_buffer;
3982 unsigned long flags;
3986 if (!cpumask_test_cpu(cpu, buffer->cpumask))
3989 dolock = rb_ok_to_lock();
3991 cpu_buffer = buffer->buffers[cpu];
3992 local_irq_save(flags);
3994 raw_spin_lock(&cpu_buffer->reader_lock);
3995 ret = rb_per_cpu_empty(cpu_buffer);
3997 raw_spin_unlock(&cpu_buffer->reader_lock);
3998 local_irq_restore(flags);
4002 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
4004 #ifdef CONFIG_RING_BUFFER_ALLOW_SWAP
4006 * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
4007 * @buffer_a: One buffer to swap with
4008 * @buffer_b: The other buffer to swap with
4010 * This function is useful for tracers that want to take a "snapshot"
4011 * of a CPU buffer and has another back up buffer lying around.
4012 * it is expected that the tracer handles the cpu buffer not being
4013 * used at the moment.
4015 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
4016 struct ring_buffer *buffer_b, int cpu)
4018 struct ring_buffer_per_cpu *cpu_buffer_a;
4019 struct ring_buffer_per_cpu *cpu_buffer_b;
4022 if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
4023 !cpumask_test_cpu(cpu, buffer_b->cpumask))
4026 cpu_buffer_a = buffer_a->buffers[cpu];
4027 cpu_buffer_b = buffer_b->buffers[cpu];
4029 /* At least make sure the two buffers are somewhat the same */
4030 if (cpu_buffer_a->nr_pages != cpu_buffer_b->nr_pages)
4035 if (ring_buffer_flags != RB_BUFFERS_ON)
4038 if (atomic_read(&buffer_a->record_disabled))
4041 if (atomic_read(&buffer_b->record_disabled))
4044 if (atomic_read(&cpu_buffer_a->record_disabled))
4047 if (atomic_read(&cpu_buffer_b->record_disabled))
4051 * We can't do a synchronize_sched here because this
4052 * function can be called in atomic context.
4053 * Normally this will be called from the same CPU as cpu.
4054 * If not it's up to the caller to protect this.
4056 atomic_inc(&cpu_buffer_a->record_disabled);
4057 atomic_inc(&cpu_buffer_b->record_disabled);
4060 if (local_read(&cpu_buffer_a->committing))
4062 if (local_read(&cpu_buffer_b->committing))
4065 buffer_a->buffers[cpu] = cpu_buffer_b;
4066 buffer_b->buffers[cpu] = cpu_buffer_a;
4068 cpu_buffer_b->buffer = buffer_a;
4069 cpu_buffer_a->buffer = buffer_b;
4074 atomic_dec(&cpu_buffer_a->record_disabled);
4075 atomic_dec(&cpu_buffer_b->record_disabled);
4079 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
4080 #endif /* CONFIG_RING_BUFFER_ALLOW_SWAP */
4083 * ring_buffer_alloc_read_page - allocate a page to read from buffer
4084 * @buffer: the buffer to allocate for.
4086 * This function is used in conjunction with ring_buffer_read_page.
4087 * When reading a full page from the ring buffer, these functions
4088 * can be used to speed up the process. The calling function should
4089 * allocate a few pages first with this function. Then when it
4090 * needs to get pages from the ring buffer, it passes the result
4091 * of this function into ring_buffer_read_page, which will swap
4092 * the page that was allocated, with the read page of the buffer.
4095 * The page allocated, or NULL on error.
4097 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer, int cpu)
4099 struct buffer_data_page *bpage;
4102 page = alloc_pages_node(cpu_to_node(cpu),
4103 GFP_KERNEL | __GFP_NORETRY, 0);
4107 bpage = page_address(page);
4109 rb_init_page(bpage);
4113 EXPORT_SYMBOL_GPL(ring_buffer_alloc_read_page);
4116 * ring_buffer_free_read_page - free an allocated read page
4117 * @buffer: the buffer the page was allocate for
4118 * @data: the page to free
4120 * Free a page allocated from ring_buffer_alloc_read_page.
4122 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
4124 free_page((unsigned long)data);
4126 EXPORT_SYMBOL_GPL(ring_buffer_free_read_page);
4129 * ring_buffer_read_page - extract a page from the ring buffer
4130 * @buffer: buffer to extract from
4131 * @data_page: the page to use allocated from ring_buffer_alloc_read_page
4132 * @len: amount to extract
4133 * @cpu: the cpu of the buffer to extract
4134 * @full: should the extraction only happen when the page is full.
4136 * This function will pull out a page from the ring buffer and consume it.
4137 * @data_page must be the address of the variable that was returned
4138 * from ring_buffer_alloc_read_page. This is because the page might be used
4139 * to swap with a page in the ring buffer.
4142 * rpage = ring_buffer_alloc_read_page(buffer);
4145 * ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
4147 * process_page(rpage, ret);
4149 * When @full is set, the function will not return true unless
4150 * the writer is off the reader page.
4152 * Note: it is up to the calling functions to handle sleeps and wakeups.
4153 * The ring buffer can be used anywhere in the kernel and can not
4154 * blindly call wake_up. The layer that uses the ring buffer must be
4155 * responsible for that.
4158 * >=0 if data has been transferred, returns the offset of consumed data.
4159 * <0 if no data has been transferred.
4161 int ring_buffer_read_page(struct ring_buffer *buffer,
4162 void **data_page, size_t len, int cpu, int full)
4164 struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
4165 struct ring_buffer_event *event;
4166 struct buffer_data_page *bpage;
4167 struct buffer_page *reader;
4168 unsigned long missed_events;
4169 unsigned long flags;
4170 unsigned int commit;
4175 if (!cpumask_test_cpu(cpu, buffer->cpumask))
4179 * If len is not big enough to hold the page header, then
4180 * we can not copy anything.
4182 if (len <= BUF_PAGE_HDR_SIZE)
4185 len -= BUF_PAGE_HDR_SIZE;
4194 raw_spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
4196 reader = rb_get_reader_page(cpu_buffer);
4200 event = rb_reader_event(cpu_buffer);
4202 read = reader->read;
4203 commit = rb_page_commit(reader);
4205 /* Check if any events were dropped */
4206 missed_events = cpu_buffer->lost_events;
4209 * If this page has been partially read or
4210 * if len is not big enough to read the rest of the page or
4211 * a writer is still on the page, then
4212 * we must copy the data from the page to the buffer.
4213 * Otherwise, we can simply swap the page with the one passed in.
4215 if (read || (len < (commit - read)) ||
4216 cpu_buffer->reader_page == cpu_buffer->commit_page) {
4217 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
4218 unsigned int rpos = read;
4219 unsigned int pos = 0;
4225 if (len > (commit - read))
4226 len = (commit - read);
4228 /* Always keep the time extend and data together */
4229 size = rb_event_ts_length(event);
4234 /* save the current timestamp, since the user will need it */
4235 save_timestamp = cpu_buffer->read_stamp;
4237 /* Need to copy one event at a time */
4239 /* We need the size of one event, because
4240 * rb_advance_reader only advances by one event,
4241 * whereas rb_event_ts_length may include the size of
4242 * one or two events.
4243 * We have already ensured there's enough space if this
4244 * is a time extend. */
4245 size = rb_event_length(event);
4246 memcpy(bpage->data + pos, rpage->data + rpos, size);
4250 rb_advance_reader(cpu_buffer);
4251 rpos = reader->read;
4257 event = rb_reader_event(cpu_buffer);
4258 /* Always keep the time extend and data together */
4259 size = rb_event_ts_length(event);
4260 } while (len >= size);
4263 local_set(&bpage->commit, pos);
4264 bpage->time_stamp = save_timestamp;
4266 /* we copied everything to the beginning */
4269 /* update the entry counter */
4270 cpu_buffer->read += rb_page_entries(reader);
4271 cpu_buffer->read_bytes += BUF_PAGE_SIZE;
4273 /* swap the pages */
4274 rb_init_page(bpage);
4275 bpage = reader->page;
4276 reader->page = *data_page;
4277 local_set(&reader->write, 0);
4278 local_set(&reader->entries, 0);
4283 * Use the real_end for the data size,
4284 * This gives us a chance to store the lost events
4287 if (reader->real_end)
4288 local_set(&bpage->commit, reader->real_end);
4292 cpu_buffer->lost_events = 0;
4294 commit = local_read(&bpage->commit);
4296 * Set a flag in the commit field if we lost events
4298 if (missed_events) {
4299 /* If there is room at the end of the page to save the
4300 * missed events, then record it there.
4302 if (BUF_PAGE_SIZE - commit >= sizeof(missed_events)) {
4303 memcpy(&bpage->data[commit], &missed_events,
4304 sizeof(missed_events));
4305 local_add(RB_MISSED_STORED, &bpage->commit);
4306 commit += sizeof(missed_events);
4308 local_add(RB_MISSED_EVENTS, &bpage->commit);
4312 * This page may be off to user land. Zero it out here.
4314 if (commit < BUF_PAGE_SIZE)
4315 memset(&bpage->data[commit], 0, BUF_PAGE_SIZE - commit);
4318 raw_spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
4323 EXPORT_SYMBOL_GPL(ring_buffer_read_page);
4325 #ifdef CONFIG_HOTPLUG_CPU
4326 static int rb_cpu_notify(struct notifier_block *self,
4327 unsigned long action, void *hcpu)
4329 struct ring_buffer *buffer =
4330 container_of(self, struct ring_buffer, cpu_notify);
4331 long cpu = (long)hcpu;
4332 int cpu_i, nr_pages_same;
4333 unsigned int nr_pages;
4336 case CPU_UP_PREPARE:
4337 case CPU_UP_PREPARE_FROZEN:
4338 if (cpumask_test_cpu(cpu, buffer->cpumask))
4343 /* check if all cpu sizes are same */
4344 for_each_buffer_cpu(buffer, cpu_i) {
4345 /* fill in the size from first enabled cpu */
4347 nr_pages = buffer->buffers[cpu_i]->nr_pages;
4348 if (nr_pages != buffer->buffers[cpu_i]->nr_pages) {
4353 /* allocate minimum pages, user can later expand it */
4356 buffer->buffers[cpu] =
4357 rb_allocate_cpu_buffer(buffer, nr_pages, cpu);
4358 if (!buffer->buffers[cpu]) {
4359 WARN(1, "failed to allocate ring buffer on CPU %ld\n",
4364 cpumask_set_cpu(cpu, buffer->cpumask);
4366 case CPU_DOWN_PREPARE:
4367 case CPU_DOWN_PREPARE_FROZEN:
4370 * If we were to free the buffer, then the user would
4371 * lose any trace that was in the buffer.