Merge git://git.kernel.org/pub/scm/linux/kernel/git/davem/net-2.6
[linux-2.6] / kernel / trace / ring_buffer.c
1 /*
2  * Generic ring buffer
3  *
4  * Copyright (C) 2008 Steven Rostedt <srostedt@redhat.com>
5  */
6 #include <linux/ring_buffer.h>
7 #include <linux/trace_clock.h>
8 #include <linux/ftrace_irq.h>
9 #include <linux/spinlock.h>
10 #include <linux/debugfs.h>
11 #include <linux/uaccess.h>
12 #include <linux/hardirq.h>
13 #include <linux/module.h>
14 #include <linux/percpu.h>
15 #include <linux/mutex.h>
16 #include <linux/init.h>
17 #include <linux/hash.h>
18 #include <linux/list.h>
19 #include <linux/cpu.h>
20 #include <linux/fs.h>
21
22 #include "trace.h"
23
24 /*
25  * The ring buffer is made up of a list of pages. A separate list of pages is
26  * allocated for each CPU. A writer may only write to a buffer that is
27  * associated with the CPU it is currently executing on.  A reader may read
28  * from any per cpu buffer.
29  *
30  * The reader is special. For each per cpu buffer, the reader has its own
31  * reader page. When a reader has read the entire reader page, this reader
32  * page is swapped with another page in the ring buffer.
33  *
34  * Now, as long as the writer is off the reader page, the reader can do what
35  * ever it wants with that page. The writer will never write to that page
36  * again (as long as it is out of the ring buffer).
37  *
38  * Here's some silly ASCII art.
39  *
40  *   +------+
41  *   |reader|          RING BUFFER
42  *   |page  |
43  *   +------+        +---+   +---+   +---+
44  *                   |   |-->|   |-->|   |
45  *                   +---+   +---+   +---+
46  *                     ^               |
47  *                     |               |
48  *                     +---------------+
49  *
50  *
51  *   +------+
52  *   |reader|          RING BUFFER
53  *   |page  |------------------v
54  *   +------+        +---+   +---+   +---+
55  *                   |   |-->|   |-->|   |
56  *                   +---+   +---+   +---+
57  *                     ^               |
58  *                     |               |
59  *                     +---------------+
60  *
61  *
62  *   +------+
63  *   |reader|          RING BUFFER
64  *   |page  |------------------v
65  *   +------+        +---+   +---+   +---+
66  *      ^            |   |-->|   |-->|   |
67  *      |            +---+   +---+   +---+
68  *      |                              |
69  *      |                              |
70  *      +------------------------------+
71  *
72  *
73  *   +------+
74  *   |buffer|          RING BUFFER
75  *   |page  |------------------v
76  *   +------+        +---+   +---+   +---+
77  *      ^            |   |   |   |-->|   |
78  *      |   New      +---+   +---+   +---+
79  *      |  Reader------^               |
80  *      |   page                       |
81  *      +------------------------------+
82  *
83  *
84  * After we make this swap, the reader can hand this page off to the splice
85  * code and be done with it. It can even allocate a new page if it needs to
86  * and swap that into the ring buffer.
87  *
88  * We will be using cmpxchg soon to make all this lockless.
89  *
90  */
91
92 /*
93  * A fast way to enable or disable all ring buffers is to
94  * call tracing_on or tracing_off. Turning off the ring buffers
95  * prevents all ring buffers from being recorded to.
96  * Turning this switch on, makes it OK to write to the
97  * ring buffer, if the ring buffer is enabled itself.
98  *
99  * There's three layers that must be on in order to write
100  * to the ring buffer.
101  *
102  * 1) This global flag must be set.
103  * 2) The ring buffer must be enabled for recording.
104  * 3) The per cpu buffer must be enabled for recording.
105  *
106  * In case of an anomaly, this global flag has a bit set that
107  * will permantly disable all ring buffers.
108  */
109
110 /*
111  * Global flag to disable all recording to ring buffers
112  *  This has two bits: ON, DISABLED
113  *
114  *  ON   DISABLED
115  * ---- ----------
116  *   0      0        : ring buffers are off
117  *   1      0        : ring buffers are on
118  *   X      1        : ring buffers are permanently disabled
119  */
120
121 enum {
122         RB_BUFFERS_ON_BIT       = 0,
123         RB_BUFFERS_DISABLED_BIT = 1,
124 };
125
126 enum {
127         RB_BUFFERS_ON           = 1 << RB_BUFFERS_ON_BIT,
128         RB_BUFFERS_DISABLED     = 1 << RB_BUFFERS_DISABLED_BIT,
129 };
130
131 static unsigned long ring_buffer_flags __read_mostly = RB_BUFFERS_ON;
132
133 #define BUF_PAGE_HDR_SIZE offsetof(struct buffer_data_page, data)
134
135 /**
136  * tracing_on - enable all tracing buffers
137  *
138  * This function enables all tracing buffers that may have been
139  * disabled with tracing_off.
140  */
141 void tracing_on(void)
142 {
143         set_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
144 }
145 EXPORT_SYMBOL_GPL(tracing_on);
146
147 /**
148  * tracing_off - turn off all tracing buffers
149  *
150  * This function stops all tracing buffers from recording data.
151  * It does not disable any overhead the tracers themselves may
152  * be causing. This function simply causes all recording to
153  * the ring buffers to fail.
154  */
155 void tracing_off(void)
156 {
157         clear_bit(RB_BUFFERS_ON_BIT, &ring_buffer_flags);
158 }
159 EXPORT_SYMBOL_GPL(tracing_off);
160
161 /**
162  * tracing_off_permanent - permanently disable ring buffers
163  *
164  * This function, once called, will disable all ring buffers
165  * permanently.
166  */
167 void tracing_off_permanent(void)
168 {
169         set_bit(RB_BUFFERS_DISABLED_BIT, &ring_buffer_flags);
170 }
171
172 /**
173  * tracing_is_on - show state of ring buffers enabled
174  */
175 int tracing_is_on(void)
176 {
177         return ring_buffer_flags == RB_BUFFERS_ON;
178 }
179 EXPORT_SYMBOL_GPL(tracing_is_on);
180
181 #include "trace.h"
182
183 #define RB_EVNT_HDR_SIZE (offsetof(struct ring_buffer_event, array))
184 #define RB_ALIGNMENT            4U
185 #define RB_MAX_SMALL_DATA       28
186
187 enum {
188         RB_LEN_TIME_EXTEND = 8,
189         RB_LEN_TIME_STAMP = 16,
190 };
191
192 static inline int rb_null_event(struct ring_buffer_event *event)
193 {
194         return event->type == RINGBUF_TYPE_PADDING && event->time_delta == 0;
195 }
196
197 static inline int rb_discarded_event(struct ring_buffer_event *event)
198 {
199         return event->type == RINGBUF_TYPE_PADDING && event->time_delta;
200 }
201
202 static void rb_event_set_padding(struct ring_buffer_event *event)
203 {
204         event->type = RINGBUF_TYPE_PADDING;
205         event->time_delta = 0;
206 }
207
208 /**
209  * ring_buffer_event_discard - discard an event in the ring buffer
210  * @buffer: the ring buffer
211  * @event: the event to discard
212  *
213  * Sometimes a event that is in the ring buffer needs to be ignored.
214  * This function lets the user discard an event in the ring buffer
215  * and then that event will not be read later.
216  *
217  * Note, it is up to the user to be careful with this, and protect
218  * against races. If the user discards an event that has been consumed
219  * it is possible that it could corrupt the ring buffer.
220  */
221 void ring_buffer_event_discard(struct ring_buffer_event *event)
222 {
223         event->type = RINGBUF_TYPE_PADDING;
224         /* time delta must be non zero */
225         if (!event->time_delta)
226                 event->time_delta = 1;
227 }
228
229 static unsigned
230 rb_event_data_length(struct ring_buffer_event *event)
231 {
232         unsigned length;
233
234         if (event->len)
235                 length = event->len * RB_ALIGNMENT;
236         else
237                 length = event->array[0];
238         return length + RB_EVNT_HDR_SIZE;
239 }
240
241 /* inline for ring buffer fast paths */
242 static unsigned
243 rb_event_length(struct ring_buffer_event *event)
244 {
245         switch (event->type) {
246         case RINGBUF_TYPE_PADDING:
247                 if (rb_null_event(event))
248                         /* undefined */
249                         return -1;
250                 return rb_event_data_length(event);
251
252         case RINGBUF_TYPE_TIME_EXTEND:
253                 return RB_LEN_TIME_EXTEND;
254
255         case RINGBUF_TYPE_TIME_STAMP:
256                 return RB_LEN_TIME_STAMP;
257
258         case RINGBUF_TYPE_DATA:
259                 return rb_event_data_length(event);
260         default:
261                 BUG();
262         }
263         /* not hit */
264         return 0;
265 }
266
267 /**
268  * ring_buffer_event_length - return the length of the event
269  * @event: the event to get the length of
270  */
271 unsigned ring_buffer_event_length(struct ring_buffer_event *event)
272 {
273         unsigned length = rb_event_length(event);
274         if (event->type != RINGBUF_TYPE_DATA)
275                 return length;
276         length -= RB_EVNT_HDR_SIZE;
277         if (length > RB_MAX_SMALL_DATA + sizeof(event->array[0]))
278                 length -= sizeof(event->array[0]);
279         return length;
280 }
281 EXPORT_SYMBOL_GPL(ring_buffer_event_length);
282
283 /* inline for ring buffer fast paths */
284 static void *
285 rb_event_data(struct ring_buffer_event *event)
286 {
287         BUG_ON(event->type != RINGBUF_TYPE_DATA);
288         /* If length is in len field, then array[0] has the data */
289         if (event->len)
290                 return (void *)&event->array[0];
291         /* Otherwise length is in array[0] and array[1] has the data */
292         return (void *)&event->array[1];
293 }
294
295 /**
296  * ring_buffer_event_data - return the data of the event
297  * @event: the event to get the data from
298  */
299 void *ring_buffer_event_data(struct ring_buffer_event *event)
300 {
301         return rb_event_data(event);
302 }
303 EXPORT_SYMBOL_GPL(ring_buffer_event_data);
304
305 #define for_each_buffer_cpu(buffer, cpu)                \
306         for_each_cpu(cpu, buffer->cpumask)
307
308 #define TS_SHIFT        27
309 #define TS_MASK         ((1ULL << TS_SHIFT) - 1)
310 #define TS_DELTA_TEST   (~TS_MASK)
311
312 struct buffer_data_page {
313         u64              time_stamp;    /* page time stamp */
314         local_t          commit;        /* write committed index */
315         unsigned char    data[];        /* data of buffer page */
316 };
317
318 struct buffer_page {
319         local_t          write;         /* index for next write */
320         unsigned         read;          /* index for next read */
321         struct list_head list;          /* list of free pages */
322         struct buffer_data_page *page;  /* Actual data page */
323 };
324
325 static void rb_init_page(struct buffer_data_page *bpage)
326 {
327         local_set(&bpage->commit, 0);
328 }
329
330 /**
331  * ring_buffer_page_len - the size of data on the page.
332  * @page: The page to read
333  *
334  * Returns the amount of data on the page, including buffer page header.
335  */
336 size_t ring_buffer_page_len(void *page)
337 {
338         return local_read(&((struct buffer_data_page *)page)->commit)
339                 + BUF_PAGE_HDR_SIZE;
340 }
341
342 /*
343  * Also stolen from mm/slob.c. Thanks to Mathieu Desnoyers for pointing
344  * this issue out.
345  */
346 static void free_buffer_page(struct buffer_page *bpage)
347 {
348         free_page((unsigned long)bpage->page);
349         kfree(bpage);
350 }
351
352 /*
353  * We need to fit the time_stamp delta into 27 bits.
354  */
355 static inline int test_time_stamp(u64 delta)
356 {
357         if (delta & TS_DELTA_TEST)
358                 return 1;
359         return 0;
360 }
361
362 #define BUF_PAGE_SIZE (PAGE_SIZE - BUF_PAGE_HDR_SIZE)
363
364 /*
365  * head_page == tail_page && head == tail then buffer is empty.
366  */
367 struct ring_buffer_per_cpu {
368         int                             cpu;
369         struct ring_buffer              *buffer;
370         spinlock_t                      reader_lock; /* serialize readers */
371         raw_spinlock_t                  lock;
372         struct lock_class_key           lock_key;
373         struct list_head                pages;
374         struct buffer_page              *head_page;     /* read from head */
375         struct buffer_page              *tail_page;     /* write to tail */
376         struct buffer_page              *commit_page;   /* committed pages */
377         struct buffer_page              *reader_page;
378         unsigned long                   overrun;
379         unsigned long                   entries;
380         u64                             write_stamp;
381         u64                             read_stamp;
382         atomic_t                        record_disabled;
383 };
384
385 struct ring_buffer {
386         unsigned                        pages;
387         unsigned                        flags;
388         int                             cpus;
389         atomic_t                        record_disabled;
390         cpumask_var_t                   cpumask;
391
392         struct mutex                    mutex;
393
394         struct ring_buffer_per_cpu      **buffers;
395
396 #ifdef CONFIG_HOTPLUG_CPU
397         struct notifier_block           cpu_notify;
398 #endif
399         u64                             (*clock)(void);
400 };
401
402 struct ring_buffer_iter {
403         struct ring_buffer_per_cpu      *cpu_buffer;
404         unsigned long                   head;
405         struct buffer_page              *head_page;
406         u64                             read_stamp;
407 };
408
409 /* buffer may be either ring_buffer or ring_buffer_per_cpu */
410 #define RB_WARN_ON(buffer, cond)                                \
411         ({                                                      \
412                 int _____ret = unlikely(cond);                  \
413                 if (_____ret) {                                 \
414                         atomic_inc(&buffer->record_disabled);   \
415                         WARN_ON(1);                             \
416                 }                                               \
417                 _____ret;                                       \
418         })
419
420 /* Up this if you want to test the TIME_EXTENTS and normalization */
421 #define DEBUG_SHIFT 0
422
423 u64 ring_buffer_time_stamp(struct ring_buffer *buffer, int cpu)
424 {
425         u64 time;
426
427         preempt_disable_notrace();
428         /* shift to debug/test normalization and TIME_EXTENTS */
429         time = buffer->clock() << DEBUG_SHIFT;
430         preempt_enable_no_resched_notrace();
431
432         return time;
433 }
434 EXPORT_SYMBOL_GPL(ring_buffer_time_stamp);
435
436 void ring_buffer_normalize_time_stamp(struct ring_buffer *buffer,
437                                       int cpu, u64 *ts)
438 {
439         /* Just stupid testing the normalize function and deltas */
440         *ts >>= DEBUG_SHIFT;
441 }
442 EXPORT_SYMBOL_GPL(ring_buffer_normalize_time_stamp);
443
444 /**
445  * check_pages - integrity check of buffer pages
446  * @cpu_buffer: CPU buffer with pages to test
447  *
448  * As a safety measure we check to make sure the data pages have not
449  * been corrupted.
450  */
451 static int rb_check_pages(struct ring_buffer_per_cpu *cpu_buffer)
452 {
453         struct list_head *head = &cpu_buffer->pages;
454         struct buffer_page *bpage, *tmp;
455
456         if (RB_WARN_ON(cpu_buffer, head->next->prev != head))
457                 return -1;
458         if (RB_WARN_ON(cpu_buffer, head->prev->next != head))
459                 return -1;
460
461         list_for_each_entry_safe(bpage, tmp, head, list) {
462                 if (RB_WARN_ON(cpu_buffer,
463                                bpage->list.next->prev != &bpage->list))
464                         return -1;
465                 if (RB_WARN_ON(cpu_buffer,
466                                bpage->list.prev->next != &bpage->list))
467                         return -1;
468         }
469
470         return 0;
471 }
472
473 static int rb_allocate_pages(struct ring_buffer_per_cpu *cpu_buffer,
474                              unsigned nr_pages)
475 {
476         struct list_head *head = &cpu_buffer->pages;
477         struct buffer_page *bpage, *tmp;
478         unsigned long addr;
479         LIST_HEAD(pages);
480         unsigned i;
481
482         for (i = 0; i < nr_pages; i++) {
483                 bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
484                                     GFP_KERNEL, cpu_to_node(cpu_buffer->cpu));
485                 if (!bpage)
486                         goto free_pages;
487                 list_add(&bpage->list, &pages);
488
489                 addr = __get_free_page(GFP_KERNEL);
490                 if (!addr)
491                         goto free_pages;
492                 bpage->page = (void *)addr;
493                 rb_init_page(bpage->page);
494         }
495
496         list_splice(&pages, head);
497
498         rb_check_pages(cpu_buffer);
499
500         return 0;
501
502  free_pages:
503         list_for_each_entry_safe(bpage, tmp, &pages, list) {
504                 list_del_init(&bpage->list);
505                 free_buffer_page(bpage);
506         }
507         return -ENOMEM;
508 }
509
510 static struct ring_buffer_per_cpu *
511 rb_allocate_cpu_buffer(struct ring_buffer *buffer, int cpu)
512 {
513         struct ring_buffer_per_cpu *cpu_buffer;
514         struct buffer_page *bpage;
515         unsigned long addr;
516         int ret;
517
518         cpu_buffer = kzalloc_node(ALIGN(sizeof(*cpu_buffer), cache_line_size()),
519                                   GFP_KERNEL, cpu_to_node(cpu));
520         if (!cpu_buffer)
521                 return NULL;
522
523         cpu_buffer->cpu = cpu;
524         cpu_buffer->buffer = buffer;
525         spin_lock_init(&cpu_buffer->reader_lock);
526         cpu_buffer->lock = (raw_spinlock_t)__RAW_SPIN_LOCK_UNLOCKED;
527         INIT_LIST_HEAD(&cpu_buffer->pages);
528
529         bpage = kzalloc_node(ALIGN(sizeof(*bpage), cache_line_size()),
530                             GFP_KERNEL, cpu_to_node(cpu));
531         if (!bpage)
532                 goto fail_free_buffer;
533
534         cpu_buffer->reader_page = bpage;
535         addr = __get_free_page(GFP_KERNEL);
536         if (!addr)
537                 goto fail_free_reader;
538         bpage->page = (void *)addr;
539         rb_init_page(bpage->page);
540
541         INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
542
543         ret = rb_allocate_pages(cpu_buffer, buffer->pages);
544         if (ret < 0)
545                 goto fail_free_reader;
546
547         cpu_buffer->head_page
548                 = list_entry(cpu_buffer->pages.next, struct buffer_page, list);
549         cpu_buffer->tail_page = cpu_buffer->commit_page = cpu_buffer->head_page;
550
551         return cpu_buffer;
552
553  fail_free_reader:
554         free_buffer_page(cpu_buffer->reader_page);
555
556  fail_free_buffer:
557         kfree(cpu_buffer);
558         return NULL;
559 }
560
561 static void rb_free_cpu_buffer(struct ring_buffer_per_cpu *cpu_buffer)
562 {
563         struct list_head *head = &cpu_buffer->pages;
564         struct buffer_page *bpage, *tmp;
565
566         free_buffer_page(cpu_buffer->reader_page);
567
568         list_for_each_entry_safe(bpage, tmp, head, list) {
569                 list_del_init(&bpage->list);
570                 free_buffer_page(bpage);
571         }
572         kfree(cpu_buffer);
573 }
574
575 /*
576  * Causes compile errors if the struct buffer_page gets bigger
577  * than the struct page.
578  */
579 extern int ring_buffer_page_too_big(void);
580
581 #ifdef CONFIG_HOTPLUG_CPU
582 static int rb_cpu_notify(struct notifier_block *self,
583                          unsigned long action, void *hcpu);
584 #endif
585
586 /**
587  * ring_buffer_alloc - allocate a new ring_buffer
588  * @size: the size in bytes per cpu that is needed.
589  * @flags: attributes to set for the ring buffer.
590  *
591  * Currently the only flag that is available is the RB_FL_OVERWRITE
592  * flag. This flag means that the buffer will overwrite old data
593  * when the buffer wraps. If this flag is not set, the buffer will
594  * drop data when the tail hits the head.
595  */
596 struct ring_buffer *ring_buffer_alloc(unsigned long size, unsigned flags)
597 {
598         struct ring_buffer *buffer;
599         int bsize;
600         int cpu;
601
602         /* Paranoid! Optimizes out when all is well */
603         if (sizeof(struct buffer_page) > sizeof(struct page))
604                 ring_buffer_page_too_big();
605
606
607         /* keep it in its own cache line */
608         buffer = kzalloc(ALIGN(sizeof(*buffer), cache_line_size()),
609                          GFP_KERNEL);
610         if (!buffer)
611                 return NULL;
612
613         if (!alloc_cpumask_var(&buffer->cpumask, GFP_KERNEL))
614                 goto fail_free_buffer;
615
616         buffer->pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
617         buffer->flags = flags;
618         buffer->clock = trace_clock_local;
619
620         /* need at least two pages */
621         if (buffer->pages == 1)
622                 buffer->pages++;
623
624         /*
625          * In case of non-hotplug cpu, if the ring-buffer is allocated
626          * in early initcall, it will not be notified of secondary cpus.
627          * In that off case, we need to allocate for all possible cpus.
628          */
629 #ifdef CONFIG_HOTPLUG_CPU
630         get_online_cpus();
631         cpumask_copy(buffer->cpumask, cpu_online_mask);
632 #else
633         cpumask_copy(buffer->cpumask, cpu_possible_mask);
634 #endif
635         buffer->cpus = nr_cpu_ids;
636
637         bsize = sizeof(void *) * nr_cpu_ids;
638         buffer->buffers = kzalloc(ALIGN(bsize, cache_line_size()),
639                                   GFP_KERNEL);
640         if (!buffer->buffers)
641                 goto fail_free_cpumask;
642
643         for_each_buffer_cpu(buffer, cpu) {
644                 buffer->buffers[cpu] =
645                         rb_allocate_cpu_buffer(buffer, cpu);
646                 if (!buffer->buffers[cpu])
647                         goto fail_free_buffers;
648         }
649
650 #ifdef CONFIG_HOTPLUG_CPU
651         buffer->cpu_notify.notifier_call = rb_cpu_notify;
652         buffer->cpu_notify.priority = 0;
653         register_cpu_notifier(&buffer->cpu_notify);
654 #endif
655
656         put_online_cpus();
657         mutex_init(&buffer->mutex);
658
659         return buffer;
660
661  fail_free_buffers:
662         for_each_buffer_cpu(buffer, cpu) {
663                 if (buffer->buffers[cpu])
664                         rb_free_cpu_buffer(buffer->buffers[cpu]);
665         }
666         kfree(buffer->buffers);
667
668  fail_free_cpumask:
669         free_cpumask_var(buffer->cpumask);
670         put_online_cpus();
671
672  fail_free_buffer:
673         kfree(buffer);
674         return NULL;
675 }
676 EXPORT_SYMBOL_GPL(ring_buffer_alloc);
677
678 /**
679  * ring_buffer_free - free a ring buffer.
680  * @buffer: the buffer to free.
681  */
682 void
683 ring_buffer_free(struct ring_buffer *buffer)
684 {
685         int cpu;
686
687         get_online_cpus();
688
689 #ifdef CONFIG_HOTPLUG_CPU
690         unregister_cpu_notifier(&buffer->cpu_notify);
691 #endif
692
693         for_each_buffer_cpu(buffer, cpu)
694                 rb_free_cpu_buffer(buffer->buffers[cpu]);
695
696         put_online_cpus();
697
698         free_cpumask_var(buffer->cpumask);
699
700         kfree(buffer);
701 }
702 EXPORT_SYMBOL_GPL(ring_buffer_free);
703
704 void ring_buffer_set_clock(struct ring_buffer *buffer,
705                            u64 (*clock)(void))
706 {
707         buffer->clock = clock;
708 }
709
710 static void rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer);
711
712 static void
713 rb_remove_pages(struct ring_buffer_per_cpu *cpu_buffer, unsigned nr_pages)
714 {
715         struct buffer_page *bpage;
716         struct list_head *p;
717         unsigned i;
718
719         atomic_inc(&cpu_buffer->record_disabled);
720         synchronize_sched();
721
722         for (i = 0; i < nr_pages; i++) {
723                 if (RB_WARN_ON(cpu_buffer, list_empty(&cpu_buffer->pages)))
724                         return;
725                 p = cpu_buffer->pages.next;
726                 bpage = list_entry(p, struct buffer_page, list);
727                 list_del_init(&bpage->list);
728                 free_buffer_page(bpage);
729         }
730         if (RB_WARN_ON(cpu_buffer, list_empty(&cpu_buffer->pages)))
731                 return;
732
733         rb_reset_cpu(cpu_buffer);
734
735         rb_check_pages(cpu_buffer);
736
737         atomic_dec(&cpu_buffer->record_disabled);
738
739 }
740
741 static void
742 rb_insert_pages(struct ring_buffer_per_cpu *cpu_buffer,
743                 struct list_head *pages, unsigned nr_pages)
744 {
745         struct buffer_page *bpage;
746         struct list_head *p;
747         unsigned i;
748
749         atomic_inc(&cpu_buffer->record_disabled);
750         synchronize_sched();
751
752         for (i = 0; i < nr_pages; i++) {
753                 if (RB_WARN_ON(cpu_buffer, list_empty(pages)))
754                         return;
755                 p = pages->next;
756                 bpage = list_entry(p, struct buffer_page, list);
757                 list_del_init(&bpage->list);
758                 list_add_tail(&bpage->list, &cpu_buffer->pages);
759         }
760         rb_reset_cpu(cpu_buffer);
761
762         rb_check_pages(cpu_buffer);
763
764         atomic_dec(&cpu_buffer->record_disabled);
765 }
766
767 /**
768  * ring_buffer_resize - resize the ring buffer
769  * @buffer: the buffer to resize.
770  * @size: the new size.
771  *
772  * The tracer is responsible for making sure that the buffer is
773  * not being used while changing the size.
774  * Note: We may be able to change the above requirement by using
775  *  RCU synchronizations.
776  *
777  * Minimum size is 2 * BUF_PAGE_SIZE.
778  *
779  * Returns -1 on failure.
780  */
781 int ring_buffer_resize(struct ring_buffer *buffer, unsigned long size)
782 {
783         struct ring_buffer_per_cpu *cpu_buffer;
784         unsigned nr_pages, rm_pages, new_pages;
785         struct buffer_page *bpage, *tmp;
786         unsigned long buffer_size;
787         unsigned long addr;
788         LIST_HEAD(pages);
789         int i, cpu;
790
791         /*
792          * Always succeed at resizing a non-existent buffer:
793          */
794         if (!buffer)
795                 return size;
796
797         size = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
798         size *= BUF_PAGE_SIZE;
799         buffer_size = buffer->pages * BUF_PAGE_SIZE;
800
801         /* we need a minimum of two pages */
802         if (size < BUF_PAGE_SIZE * 2)
803                 size = BUF_PAGE_SIZE * 2;
804
805         if (size == buffer_size)
806                 return size;
807
808         mutex_lock(&buffer->mutex);
809         get_online_cpus();
810
811         nr_pages = DIV_ROUND_UP(size, BUF_PAGE_SIZE);
812
813         if (size < buffer_size) {
814
815                 /* easy case, just free pages */
816                 if (RB_WARN_ON(buffer, nr_pages >= buffer->pages))
817                         goto out_fail;
818
819                 rm_pages = buffer->pages - nr_pages;
820
821                 for_each_buffer_cpu(buffer, cpu) {
822                         cpu_buffer = buffer->buffers[cpu];
823                         rb_remove_pages(cpu_buffer, rm_pages);
824                 }
825                 goto out;
826         }
827
828         /*
829          * This is a bit more difficult. We only want to add pages
830          * when we can allocate enough for all CPUs. We do this
831          * by allocating all the pages and storing them on a local
832          * link list. If we succeed in our allocation, then we
833          * add these pages to the cpu_buffers. Otherwise we just free
834          * them all and return -ENOMEM;
835          */
836         if (RB_WARN_ON(buffer, nr_pages <= buffer->pages))
837                 goto out_fail;
838
839         new_pages = nr_pages - buffer->pages;
840
841         for_each_buffer_cpu(buffer, cpu) {
842                 for (i = 0; i < new_pages; i++) {
843                         bpage = kzalloc_node(ALIGN(sizeof(*bpage),
844                                                   cache_line_size()),
845                                             GFP_KERNEL, cpu_to_node(cpu));
846                         if (!bpage)
847                                 goto free_pages;
848                         list_add(&bpage->list, &pages);
849                         addr = __get_free_page(GFP_KERNEL);
850                         if (!addr)
851                                 goto free_pages;
852                         bpage->page = (void *)addr;
853                         rb_init_page(bpage->page);
854                 }
855         }
856
857         for_each_buffer_cpu(buffer, cpu) {
858                 cpu_buffer = buffer->buffers[cpu];
859                 rb_insert_pages(cpu_buffer, &pages, new_pages);
860         }
861
862         if (RB_WARN_ON(buffer, !list_empty(&pages)))
863                 goto out_fail;
864
865  out:
866         buffer->pages = nr_pages;
867         put_online_cpus();
868         mutex_unlock(&buffer->mutex);
869
870         return size;
871
872  free_pages:
873         list_for_each_entry_safe(bpage, tmp, &pages, list) {
874                 list_del_init(&bpage->list);
875                 free_buffer_page(bpage);
876         }
877         put_online_cpus();
878         mutex_unlock(&buffer->mutex);
879         return -ENOMEM;
880
881         /*
882          * Something went totally wrong, and we are too paranoid
883          * to even clean up the mess.
884          */
885  out_fail:
886         put_online_cpus();
887         mutex_unlock(&buffer->mutex);
888         return -1;
889 }
890 EXPORT_SYMBOL_GPL(ring_buffer_resize);
891
892 static inline void *
893 __rb_data_page_index(struct buffer_data_page *bpage, unsigned index)
894 {
895         return bpage->data + index;
896 }
897
898 static inline void *__rb_page_index(struct buffer_page *bpage, unsigned index)
899 {
900         return bpage->page->data + index;
901 }
902
903 static inline struct ring_buffer_event *
904 rb_reader_event(struct ring_buffer_per_cpu *cpu_buffer)
905 {
906         return __rb_page_index(cpu_buffer->reader_page,
907                                cpu_buffer->reader_page->read);
908 }
909
910 static inline struct ring_buffer_event *
911 rb_head_event(struct ring_buffer_per_cpu *cpu_buffer)
912 {
913         return __rb_page_index(cpu_buffer->head_page,
914                                cpu_buffer->head_page->read);
915 }
916
917 static inline struct ring_buffer_event *
918 rb_iter_head_event(struct ring_buffer_iter *iter)
919 {
920         return __rb_page_index(iter->head_page, iter->head);
921 }
922
923 static inline unsigned rb_page_write(struct buffer_page *bpage)
924 {
925         return local_read(&bpage->write);
926 }
927
928 static inline unsigned rb_page_commit(struct buffer_page *bpage)
929 {
930         return local_read(&bpage->page->commit);
931 }
932
933 /* Size is determined by what has been commited */
934 static inline unsigned rb_page_size(struct buffer_page *bpage)
935 {
936         return rb_page_commit(bpage);
937 }
938
939 static inline unsigned
940 rb_commit_index(struct ring_buffer_per_cpu *cpu_buffer)
941 {
942         return rb_page_commit(cpu_buffer->commit_page);
943 }
944
945 static inline unsigned rb_head_size(struct ring_buffer_per_cpu *cpu_buffer)
946 {
947         return rb_page_commit(cpu_buffer->head_page);
948 }
949
950 /*
951  * When the tail hits the head and the buffer is in overwrite mode,
952  * the head jumps to the next page and all content on the previous
953  * page is discarded. But before doing so, we update the overrun
954  * variable of the buffer.
955  */
956 static void rb_update_overflow(struct ring_buffer_per_cpu *cpu_buffer)
957 {
958         struct ring_buffer_event *event;
959         unsigned long head;
960
961         for (head = 0; head < rb_head_size(cpu_buffer);
962              head += rb_event_length(event)) {
963
964                 event = __rb_page_index(cpu_buffer->head_page, head);
965                 if (RB_WARN_ON(cpu_buffer, rb_null_event(event)))
966                         return;
967                 /* Only count data entries */
968                 if (event->type != RINGBUF_TYPE_DATA)
969                         continue;
970                 cpu_buffer->overrun++;
971                 cpu_buffer->entries--;
972         }
973 }
974
975 static inline void rb_inc_page(struct ring_buffer_per_cpu *cpu_buffer,
976                                struct buffer_page **bpage)
977 {
978         struct list_head *p = (*bpage)->list.next;
979
980         if (p == &cpu_buffer->pages)
981                 p = p->next;
982
983         *bpage = list_entry(p, struct buffer_page, list);
984 }
985
986 static inline unsigned
987 rb_event_index(struct ring_buffer_event *event)
988 {
989         unsigned long addr = (unsigned long)event;
990
991         return (addr & ~PAGE_MASK) - (PAGE_SIZE - BUF_PAGE_SIZE);
992 }
993
994 static int
995 rb_is_commit(struct ring_buffer_per_cpu *cpu_buffer,
996              struct ring_buffer_event *event)
997 {
998         unsigned long addr = (unsigned long)event;
999         unsigned long index;
1000
1001         index = rb_event_index(event);
1002         addr &= PAGE_MASK;
1003
1004         return cpu_buffer->commit_page->page == (void *)addr &&
1005                 rb_commit_index(cpu_buffer) == index;
1006 }
1007
1008 static void
1009 rb_set_commit_event(struct ring_buffer_per_cpu *cpu_buffer,
1010                     struct ring_buffer_event *event)
1011 {
1012         unsigned long addr = (unsigned long)event;
1013         unsigned long index;
1014
1015         index = rb_event_index(event);
1016         addr &= PAGE_MASK;
1017
1018         while (cpu_buffer->commit_page->page != (void *)addr) {
1019                 if (RB_WARN_ON(cpu_buffer,
1020                           cpu_buffer->commit_page == cpu_buffer->tail_page))
1021                         return;
1022                 cpu_buffer->commit_page->page->commit =
1023                         cpu_buffer->commit_page->write;
1024                 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1025                 cpu_buffer->write_stamp =
1026                         cpu_buffer->commit_page->page->time_stamp;
1027         }
1028
1029         /* Now set the commit to the event's index */
1030         local_set(&cpu_buffer->commit_page->page->commit, index);
1031 }
1032
1033 static void
1034 rb_set_commit_to_write(struct ring_buffer_per_cpu *cpu_buffer)
1035 {
1036         /*
1037          * We only race with interrupts and NMIs on this CPU.
1038          * If we own the commit event, then we can commit
1039          * all others that interrupted us, since the interruptions
1040          * are in stack format (they finish before they come
1041          * back to us). This allows us to do a simple loop to
1042          * assign the commit to the tail.
1043          */
1044  again:
1045         while (cpu_buffer->commit_page != cpu_buffer->tail_page) {
1046                 cpu_buffer->commit_page->page->commit =
1047                         cpu_buffer->commit_page->write;
1048                 rb_inc_page(cpu_buffer, &cpu_buffer->commit_page);
1049                 cpu_buffer->write_stamp =
1050                         cpu_buffer->commit_page->page->time_stamp;
1051                 /* add barrier to keep gcc from optimizing too much */
1052                 barrier();
1053         }
1054         while (rb_commit_index(cpu_buffer) !=
1055                rb_page_write(cpu_buffer->commit_page)) {
1056                 cpu_buffer->commit_page->page->commit =
1057                         cpu_buffer->commit_page->write;
1058                 barrier();
1059         }
1060
1061         /* again, keep gcc from optimizing */
1062         barrier();
1063
1064         /*
1065          * If an interrupt came in just after the first while loop
1066          * and pushed the tail page forward, we will be left with
1067          * a dangling commit that will never go forward.
1068          */
1069         if (unlikely(cpu_buffer->commit_page != cpu_buffer->tail_page))
1070                 goto again;
1071 }
1072
1073 static void rb_reset_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1074 {
1075         cpu_buffer->read_stamp = cpu_buffer->reader_page->page->time_stamp;
1076         cpu_buffer->reader_page->read = 0;
1077 }
1078
1079 static void rb_inc_iter(struct ring_buffer_iter *iter)
1080 {
1081         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1082
1083         /*
1084          * The iterator could be on the reader page (it starts there).
1085          * But the head could have moved, since the reader was
1086          * found. Check for this case and assign the iterator
1087          * to the head page instead of next.
1088          */
1089         if (iter->head_page == cpu_buffer->reader_page)
1090                 iter->head_page = cpu_buffer->head_page;
1091         else
1092                 rb_inc_page(cpu_buffer, &iter->head_page);
1093
1094         iter->read_stamp = iter->head_page->page->time_stamp;
1095         iter->head = 0;
1096 }
1097
1098 /**
1099  * ring_buffer_update_event - update event type and data
1100  * @event: the even to update
1101  * @type: the type of event
1102  * @length: the size of the event field in the ring buffer
1103  *
1104  * Update the type and data fields of the event. The length
1105  * is the actual size that is written to the ring buffer,
1106  * and with this, we can determine what to place into the
1107  * data field.
1108  */
1109 static void
1110 rb_update_event(struct ring_buffer_event *event,
1111                          unsigned type, unsigned length)
1112 {
1113         event->type = type;
1114
1115         switch (type) {
1116
1117         case RINGBUF_TYPE_PADDING:
1118                 break;
1119
1120         case RINGBUF_TYPE_TIME_EXTEND:
1121                 event->len = DIV_ROUND_UP(RB_LEN_TIME_EXTEND, RB_ALIGNMENT);
1122                 break;
1123
1124         case RINGBUF_TYPE_TIME_STAMP:
1125                 event->len = DIV_ROUND_UP(RB_LEN_TIME_STAMP, RB_ALIGNMENT);
1126                 break;
1127
1128         case RINGBUF_TYPE_DATA:
1129                 length -= RB_EVNT_HDR_SIZE;
1130                 if (length > RB_MAX_SMALL_DATA) {
1131                         event->len = 0;
1132                         event->array[0] = length;
1133                 } else
1134                         event->len = DIV_ROUND_UP(length, RB_ALIGNMENT);
1135                 break;
1136         default:
1137                 BUG();
1138         }
1139 }
1140
1141 static unsigned rb_calculate_event_length(unsigned length)
1142 {
1143         struct ring_buffer_event event; /* Used only for sizeof array */
1144
1145         /* zero length can cause confusions */
1146         if (!length)
1147                 length = 1;
1148
1149         if (length > RB_MAX_SMALL_DATA)
1150                 length += sizeof(event.array[0]);
1151
1152         length += RB_EVNT_HDR_SIZE;
1153         length = ALIGN(length, RB_ALIGNMENT);
1154
1155         return length;
1156 }
1157
1158 static struct ring_buffer_event *
1159 __rb_reserve_next(struct ring_buffer_per_cpu *cpu_buffer,
1160                   unsigned type, unsigned long length, u64 *ts)
1161 {
1162         struct buffer_page *tail_page, *head_page, *reader_page, *commit_page;
1163         unsigned long tail, write;
1164         struct ring_buffer *buffer = cpu_buffer->buffer;
1165         struct ring_buffer_event *event;
1166         unsigned long flags;
1167         bool lock_taken = false;
1168
1169         commit_page = cpu_buffer->commit_page;
1170         /* we just need to protect against interrupts */
1171         barrier();
1172         tail_page = cpu_buffer->tail_page;
1173         write = local_add_return(length, &tail_page->write);
1174         tail = write - length;
1175
1176         /* See if we shot pass the end of this buffer page */
1177         if (write > BUF_PAGE_SIZE) {
1178                 struct buffer_page *next_page = tail_page;
1179
1180                 local_irq_save(flags);
1181                 /*
1182                  * Since the write to the buffer is still not
1183                  * fully lockless, we must be careful with NMIs.
1184                  * The locks in the writers are taken when a write
1185                  * crosses to a new page. The locks protect against
1186                  * races with the readers (this will soon be fixed
1187                  * with a lockless solution).
1188                  *
1189                  * Because we can not protect against NMIs, and we
1190                  * want to keep traces reentrant, we need to manage
1191                  * what happens when we are in an NMI.
1192                  *
1193                  * NMIs can happen after we take the lock.
1194                  * If we are in an NMI, only take the lock
1195                  * if it is not already taken. Otherwise
1196                  * simply fail.
1197                  */
1198                 if (unlikely(in_nmi())) {
1199                         if (!__raw_spin_trylock(&cpu_buffer->lock))
1200                                 goto out_reset;
1201                 } else
1202                         __raw_spin_lock(&cpu_buffer->lock);
1203
1204                 lock_taken = true;
1205
1206                 rb_inc_page(cpu_buffer, &next_page);
1207
1208                 head_page = cpu_buffer->head_page;
1209                 reader_page = cpu_buffer->reader_page;
1210
1211                 /* we grabbed the lock before incrementing */
1212                 if (RB_WARN_ON(cpu_buffer, next_page == reader_page))
1213                         goto out_reset;
1214
1215                 /*
1216                  * If for some reason, we had an interrupt storm that made
1217                  * it all the way around the buffer, bail, and warn
1218                  * about it.
1219                  */
1220                 if (unlikely(next_page == commit_page)) {
1221                         WARN_ON_ONCE(1);
1222                         goto out_reset;
1223                 }
1224
1225                 if (next_page == head_page) {
1226                         if (!(buffer->flags & RB_FL_OVERWRITE))
1227                                 goto out_reset;
1228
1229                         /* tail_page has not moved yet? */
1230                         if (tail_page == cpu_buffer->tail_page) {
1231                                 /* count overflows */
1232                                 rb_update_overflow(cpu_buffer);
1233
1234                                 rb_inc_page(cpu_buffer, &head_page);
1235                                 cpu_buffer->head_page = head_page;
1236                                 cpu_buffer->head_page->read = 0;
1237                         }
1238                 }
1239
1240                 /*
1241                  * If the tail page is still the same as what we think
1242                  * it is, then it is up to us to update the tail
1243                  * pointer.
1244                  */
1245                 if (tail_page == cpu_buffer->tail_page) {
1246                         local_set(&next_page->write, 0);
1247                         local_set(&next_page->page->commit, 0);
1248                         cpu_buffer->tail_page = next_page;
1249
1250                         /* reread the time stamp */
1251                         *ts = ring_buffer_time_stamp(buffer, cpu_buffer->cpu);
1252                         cpu_buffer->tail_page->page->time_stamp = *ts;
1253                 }
1254
1255                 /*
1256                  * The actual tail page has moved forward.
1257                  */
1258                 if (tail < BUF_PAGE_SIZE) {
1259                         /* Mark the rest of the page with padding */
1260                         event = __rb_page_index(tail_page, tail);
1261                         rb_event_set_padding(event);
1262                 }
1263
1264                 if (tail <= BUF_PAGE_SIZE)
1265                         /* Set the write back to the previous setting */
1266                         local_set(&tail_page->write, tail);
1267
1268                 /*
1269                  * If this was a commit entry that failed,
1270                  * increment that too
1271                  */
1272                 if (tail_page == cpu_buffer->commit_page &&
1273                     tail == rb_commit_index(cpu_buffer)) {
1274                         rb_set_commit_to_write(cpu_buffer);
1275                 }
1276
1277                 __raw_spin_unlock(&cpu_buffer->lock);
1278                 local_irq_restore(flags);
1279
1280                 /* fail and let the caller try again */
1281                 return ERR_PTR(-EAGAIN);
1282         }
1283
1284         /* We reserved something on the buffer */
1285
1286         if (RB_WARN_ON(cpu_buffer, write > BUF_PAGE_SIZE))
1287                 return NULL;
1288
1289         event = __rb_page_index(tail_page, tail);
1290         rb_update_event(event, type, length);
1291
1292         /*
1293          * If this is a commit and the tail is zero, then update
1294          * this page's time stamp.
1295          */
1296         if (!tail && rb_is_commit(cpu_buffer, event))
1297                 cpu_buffer->commit_page->page->time_stamp = *ts;
1298
1299         return event;
1300
1301  out_reset:
1302         /* reset write */
1303         if (tail <= BUF_PAGE_SIZE)
1304                 local_set(&tail_page->write, tail);
1305
1306         if (likely(lock_taken))
1307                 __raw_spin_unlock(&cpu_buffer->lock);
1308         local_irq_restore(flags);
1309         return NULL;
1310 }
1311
1312 static int
1313 rb_add_time_stamp(struct ring_buffer_per_cpu *cpu_buffer,
1314                   u64 *ts, u64 *delta)
1315 {
1316         struct ring_buffer_event *event;
1317         static int once;
1318         int ret;
1319
1320         if (unlikely(*delta > (1ULL << 59) && !once++)) {
1321                 printk(KERN_WARNING "Delta way too big! %llu"
1322                        " ts=%llu write stamp = %llu\n",
1323                        (unsigned long long)*delta,
1324                        (unsigned long long)*ts,
1325                        (unsigned long long)cpu_buffer->write_stamp);
1326                 WARN_ON(1);
1327         }
1328
1329         /*
1330          * The delta is too big, we to add a
1331          * new timestamp.
1332          */
1333         event = __rb_reserve_next(cpu_buffer,
1334                                   RINGBUF_TYPE_TIME_EXTEND,
1335                                   RB_LEN_TIME_EXTEND,
1336                                   ts);
1337         if (!event)
1338                 return -EBUSY;
1339
1340         if (PTR_ERR(event) == -EAGAIN)
1341                 return -EAGAIN;
1342
1343         /* Only a commited time event can update the write stamp */
1344         if (rb_is_commit(cpu_buffer, event)) {
1345                 /*
1346                  * If this is the first on the page, then we need to
1347                  * update the page itself, and just put in a zero.
1348                  */
1349                 if (rb_event_index(event)) {
1350                         event->time_delta = *delta & TS_MASK;
1351                         event->array[0] = *delta >> TS_SHIFT;
1352                 } else {
1353                         cpu_buffer->commit_page->page->time_stamp = *ts;
1354                         event->time_delta = 0;
1355                         event->array[0] = 0;
1356                 }
1357                 cpu_buffer->write_stamp = *ts;
1358                 /* let the caller know this was the commit */
1359                 ret = 1;
1360         } else {
1361                 /* Darn, this is just wasted space */
1362                 event->time_delta = 0;
1363                 event->array[0] = 0;
1364                 ret = 0;
1365         }
1366
1367         *delta = 0;
1368
1369         return ret;
1370 }
1371
1372 static struct ring_buffer_event *
1373 rb_reserve_next_event(struct ring_buffer_per_cpu *cpu_buffer,
1374                       unsigned type, unsigned long length)
1375 {
1376         struct ring_buffer_event *event;
1377         u64 ts, delta;
1378         int commit = 0;
1379         int nr_loops = 0;
1380
1381  again:
1382         /*
1383          * We allow for interrupts to reenter here and do a trace.
1384          * If one does, it will cause this original code to loop
1385          * back here. Even with heavy interrupts happening, this
1386          * should only happen a few times in a row. If this happens
1387          * 1000 times in a row, there must be either an interrupt
1388          * storm or we have something buggy.
1389          * Bail!
1390          */
1391         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 1000))
1392                 return NULL;
1393
1394         ts = ring_buffer_time_stamp(cpu_buffer->buffer, cpu_buffer->cpu);
1395
1396         /*
1397          * Only the first commit can update the timestamp.
1398          * Yes there is a race here. If an interrupt comes in
1399          * just after the conditional and it traces too, then it
1400          * will also check the deltas. More than one timestamp may
1401          * also be made. But only the entry that did the actual
1402          * commit will be something other than zero.
1403          */
1404         if (cpu_buffer->tail_page == cpu_buffer->commit_page &&
1405             rb_page_write(cpu_buffer->tail_page) ==
1406             rb_commit_index(cpu_buffer)) {
1407
1408                 delta = ts - cpu_buffer->write_stamp;
1409
1410                 /* make sure this delta is calculated here */
1411                 barrier();
1412
1413                 /* Did the write stamp get updated already? */
1414                 if (unlikely(ts < cpu_buffer->write_stamp))
1415                         delta = 0;
1416
1417                 if (test_time_stamp(delta)) {
1418
1419                         commit = rb_add_time_stamp(cpu_buffer, &ts, &delta);
1420
1421                         if (commit == -EBUSY)
1422                                 return NULL;
1423
1424                         if (commit == -EAGAIN)
1425                                 goto again;
1426
1427                         RB_WARN_ON(cpu_buffer, commit < 0);
1428                 }
1429         } else
1430                 /* Non commits have zero deltas */
1431                 delta = 0;
1432
1433         event = __rb_reserve_next(cpu_buffer, type, length, &ts);
1434         if (PTR_ERR(event) == -EAGAIN)
1435                 goto again;
1436
1437         if (!event) {
1438                 if (unlikely(commit))
1439                         /*
1440                          * Ouch! We needed a timestamp and it was commited. But
1441                          * we didn't get our event reserved.
1442                          */
1443                         rb_set_commit_to_write(cpu_buffer);
1444                 return NULL;
1445         }
1446
1447         /*
1448          * If the timestamp was commited, make the commit our entry
1449          * now so that we will update it when needed.
1450          */
1451         if (commit)
1452                 rb_set_commit_event(cpu_buffer, event);
1453         else if (!rb_is_commit(cpu_buffer, event))
1454                 delta = 0;
1455
1456         event->time_delta = delta;
1457
1458         return event;
1459 }
1460
1461 static DEFINE_PER_CPU(int, rb_need_resched);
1462
1463 /**
1464  * ring_buffer_lock_reserve - reserve a part of the buffer
1465  * @buffer: the ring buffer to reserve from
1466  * @length: the length of the data to reserve (excluding event header)
1467  *
1468  * Returns a reseverd event on the ring buffer to copy directly to.
1469  * The user of this interface will need to get the body to write into
1470  * and can use the ring_buffer_event_data() interface.
1471  *
1472  * The length is the length of the data needed, not the event length
1473  * which also includes the event header.
1474  *
1475  * Must be paired with ring_buffer_unlock_commit, unless NULL is returned.
1476  * If NULL is returned, then nothing has been allocated or locked.
1477  */
1478 struct ring_buffer_event *
1479 ring_buffer_lock_reserve(struct ring_buffer *buffer, unsigned long length)
1480 {
1481         struct ring_buffer_per_cpu *cpu_buffer;
1482         struct ring_buffer_event *event;
1483         int cpu, resched;
1484
1485         if (ring_buffer_flags != RB_BUFFERS_ON)
1486                 return NULL;
1487
1488         if (atomic_read(&buffer->record_disabled))
1489                 return NULL;
1490
1491         /* If we are tracing schedule, we don't want to recurse */
1492         resched = ftrace_preempt_disable();
1493
1494         cpu = raw_smp_processor_id();
1495
1496         if (!cpumask_test_cpu(cpu, buffer->cpumask))
1497                 goto out;
1498
1499         cpu_buffer = buffer->buffers[cpu];
1500
1501         if (atomic_read(&cpu_buffer->record_disabled))
1502                 goto out;
1503
1504         length = rb_calculate_event_length(length);
1505         if (length > BUF_PAGE_SIZE)
1506                 goto out;
1507
1508         event = rb_reserve_next_event(cpu_buffer, RINGBUF_TYPE_DATA, length);
1509         if (!event)
1510                 goto out;
1511
1512         /*
1513          * Need to store resched state on this cpu.
1514          * Only the first needs to.
1515          */
1516
1517         if (preempt_count() == 1)
1518                 per_cpu(rb_need_resched, cpu) = resched;
1519
1520         return event;
1521
1522  out:
1523         ftrace_preempt_enable(resched);
1524         return NULL;
1525 }
1526 EXPORT_SYMBOL_GPL(ring_buffer_lock_reserve);
1527
1528 static void rb_commit(struct ring_buffer_per_cpu *cpu_buffer,
1529                       struct ring_buffer_event *event)
1530 {
1531         cpu_buffer->entries++;
1532
1533         /* Only process further if we own the commit */
1534         if (!rb_is_commit(cpu_buffer, event))
1535                 return;
1536
1537         cpu_buffer->write_stamp += event->time_delta;
1538
1539         rb_set_commit_to_write(cpu_buffer);
1540 }
1541
1542 /**
1543  * ring_buffer_unlock_commit - commit a reserved
1544  * @buffer: The buffer to commit to
1545  * @event: The event pointer to commit.
1546  *
1547  * This commits the data to the ring buffer, and releases any locks held.
1548  *
1549  * Must be paired with ring_buffer_lock_reserve.
1550  */
1551 int ring_buffer_unlock_commit(struct ring_buffer *buffer,
1552                               struct ring_buffer_event *event)
1553 {
1554         struct ring_buffer_per_cpu *cpu_buffer;
1555         int cpu = raw_smp_processor_id();
1556
1557         cpu_buffer = buffer->buffers[cpu];
1558
1559         rb_commit(cpu_buffer, event);
1560
1561         /*
1562          * Only the last preempt count needs to restore preemption.
1563          */
1564         if (preempt_count() == 1)
1565                 ftrace_preempt_enable(per_cpu(rb_need_resched, cpu));
1566         else
1567                 preempt_enable_no_resched_notrace();
1568
1569         return 0;
1570 }
1571 EXPORT_SYMBOL_GPL(ring_buffer_unlock_commit);
1572
1573 /**
1574  * ring_buffer_write - write data to the buffer without reserving
1575  * @buffer: The ring buffer to write to.
1576  * @length: The length of the data being written (excluding the event header)
1577  * @data: The data to write to the buffer.
1578  *
1579  * This is like ring_buffer_lock_reserve and ring_buffer_unlock_commit as
1580  * one function. If you already have the data to write to the buffer, it
1581  * may be easier to simply call this function.
1582  *
1583  * Note, like ring_buffer_lock_reserve, the length is the length of the data
1584  * and not the length of the event which would hold the header.
1585  */
1586 int ring_buffer_write(struct ring_buffer *buffer,
1587                         unsigned long length,
1588                         void *data)
1589 {
1590         struct ring_buffer_per_cpu *cpu_buffer;
1591         struct ring_buffer_event *event;
1592         unsigned long event_length;
1593         void *body;
1594         int ret = -EBUSY;
1595         int cpu, resched;
1596
1597         if (ring_buffer_flags != RB_BUFFERS_ON)
1598                 return -EBUSY;
1599
1600         if (atomic_read(&buffer->record_disabled))
1601                 return -EBUSY;
1602
1603         resched = ftrace_preempt_disable();
1604
1605         cpu = raw_smp_processor_id();
1606
1607         if (!cpumask_test_cpu(cpu, buffer->cpumask))
1608                 goto out;
1609
1610         cpu_buffer = buffer->buffers[cpu];
1611
1612         if (atomic_read(&cpu_buffer->record_disabled))
1613                 goto out;
1614
1615         event_length = rb_calculate_event_length(length);
1616         event = rb_reserve_next_event(cpu_buffer,
1617                                       RINGBUF_TYPE_DATA, event_length);
1618         if (!event)
1619                 goto out;
1620
1621         body = rb_event_data(event);
1622
1623         memcpy(body, data, length);
1624
1625         rb_commit(cpu_buffer, event);
1626
1627         ret = 0;
1628  out:
1629         ftrace_preempt_enable(resched);
1630
1631         return ret;
1632 }
1633 EXPORT_SYMBOL_GPL(ring_buffer_write);
1634
1635 static int rb_per_cpu_empty(struct ring_buffer_per_cpu *cpu_buffer)
1636 {
1637         struct buffer_page *reader = cpu_buffer->reader_page;
1638         struct buffer_page *head = cpu_buffer->head_page;
1639         struct buffer_page *commit = cpu_buffer->commit_page;
1640
1641         return reader->read == rb_page_commit(reader) &&
1642                 (commit == reader ||
1643                  (commit == head &&
1644                   head->read == rb_page_commit(commit)));
1645 }
1646
1647 /**
1648  * ring_buffer_record_disable - stop all writes into the buffer
1649  * @buffer: The ring buffer to stop writes to.
1650  *
1651  * This prevents all writes to the buffer. Any attempt to write
1652  * to the buffer after this will fail and return NULL.
1653  *
1654  * The caller should call synchronize_sched() after this.
1655  */
1656 void ring_buffer_record_disable(struct ring_buffer *buffer)
1657 {
1658         atomic_inc(&buffer->record_disabled);
1659 }
1660 EXPORT_SYMBOL_GPL(ring_buffer_record_disable);
1661
1662 /**
1663  * ring_buffer_record_enable - enable writes to the buffer
1664  * @buffer: The ring buffer to enable writes
1665  *
1666  * Note, multiple disables will need the same number of enables
1667  * to truely enable the writing (much like preempt_disable).
1668  */
1669 void ring_buffer_record_enable(struct ring_buffer *buffer)
1670 {
1671         atomic_dec(&buffer->record_disabled);
1672 }
1673 EXPORT_SYMBOL_GPL(ring_buffer_record_enable);
1674
1675 /**
1676  * ring_buffer_record_disable_cpu - stop all writes into the cpu_buffer
1677  * @buffer: The ring buffer to stop writes to.
1678  * @cpu: The CPU buffer to stop
1679  *
1680  * This prevents all writes to the buffer. Any attempt to write
1681  * to the buffer after this will fail and return NULL.
1682  *
1683  * The caller should call synchronize_sched() after this.
1684  */
1685 void ring_buffer_record_disable_cpu(struct ring_buffer *buffer, int cpu)
1686 {
1687         struct ring_buffer_per_cpu *cpu_buffer;
1688
1689         if (!cpumask_test_cpu(cpu, buffer->cpumask))
1690                 return;
1691
1692         cpu_buffer = buffer->buffers[cpu];
1693         atomic_inc(&cpu_buffer->record_disabled);
1694 }
1695 EXPORT_SYMBOL_GPL(ring_buffer_record_disable_cpu);
1696
1697 /**
1698  * ring_buffer_record_enable_cpu - enable writes to the buffer
1699  * @buffer: The ring buffer to enable writes
1700  * @cpu: The CPU to enable.
1701  *
1702  * Note, multiple disables will need the same number of enables
1703  * to truely enable the writing (much like preempt_disable).
1704  */
1705 void ring_buffer_record_enable_cpu(struct ring_buffer *buffer, int cpu)
1706 {
1707         struct ring_buffer_per_cpu *cpu_buffer;
1708
1709         if (!cpumask_test_cpu(cpu, buffer->cpumask))
1710                 return;
1711
1712         cpu_buffer = buffer->buffers[cpu];
1713         atomic_dec(&cpu_buffer->record_disabled);
1714 }
1715 EXPORT_SYMBOL_GPL(ring_buffer_record_enable_cpu);
1716
1717 /**
1718  * ring_buffer_entries_cpu - get the number of entries in a cpu buffer
1719  * @buffer: The ring buffer
1720  * @cpu: The per CPU buffer to get the entries from.
1721  */
1722 unsigned long ring_buffer_entries_cpu(struct ring_buffer *buffer, int cpu)
1723 {
1724         struct ring_buffer_per_cpu *cpu_buffer;
1725         unsigned long ret;
1726
1727         if (!cpumask_test_cpu(cpu, buffer->cpumask))
1728                 return 0;
1729
1730         cpu_buffer = buffer->buffers[cpu];
1731         ret = cpu_buffer->entries;
1732
1733         return ret;
1734 }
1735 EXPORT_SYMBOL_GPL(ring_buffer_entries_cpu);
1736
1737 /**
1738  * ring_buffer_overrun_cpu - get the number of overruns in a cpu_buffer
1739  * @buffer: The ring buffer
1740  * @cpu: The per CPU buffer to get the number of overruns from
1741  */
1742 unsigned long ring_buffer_overrun_cpu(struct ring_buffer *buffer, int cpu)
1743 {
1744         struct ring_buffer_per_cpu *cpu_buffer;
1745         unsigned long ret;
1746
1747         if (!cpumask_test_cpu(cpu, buffer->cpumask))
1748                 return 0;
1749
1750         cpu_buffer = buffer->buffers[cpu];
1751         ret = cpu_buffer->overrun;
1752
1753         return ret;
1754 }
1755 EXPORT_SYMBOL_GPL(ring_buffer_overrun_cpu);
1756
1757 /**
1758  * ring_buffer_entries - get the number of entries in a buffer
1759  * @buffer: The ring buffer
1760  *
1761  * Returns the total number of entries in the ring buffer
1762  * (all CPU entries)
1763  */
1764 unsigned long ring_buffer_entries(struct ring_buffer *buffer)
1765 {
1766         struct ring_buffer_per_cpu *cpu_buffer;
1767         unsigned long entries = 0;
1768         int cpu;
1769
1770         /* if you care about this being correct, lock the buffer */
1771         for_each_buffer_cpu(buffer, cpu) {
1772                 cpu_buffer = buffer->buffers[cpu];
1773                 entries += cpu_buffer->entries;
1774         }
1775
1776         return entries;
1777 }
1778 EXPORT_SYMBOL_GPL(ring_buffer_entries);
1779
1780 /**
1781  * ring_buffer_overrun_cpu - get the number of overruns in buffer
1782  * @buffer: The ring buffer
1783  *
1784  * Returns the total number of overruns in the ring buffer
1785  * (all CPU entries)
1786  */
1787 unsigned long ring_buffer_overruns(struct ring_buffer *buffer)
1788 {
1789         struct ring_buffer_per_cpu *cpu_buffer;
1790         unsigned long overruns = 0;
1791         int cpu;
1792
1793         /* if you care about this being correct, lock the buffer */
1794         for_each_buffer_cpu(buffer, cpu) {
1795                 cpu_buffer = buffer->buffers[cpu];
1796                 overruns += cpu_buffer->overrun;
1797         }
1798
1799         return overruns;
1800 }
1801 EXPORT_SYMBOL_GPL(ring_buffer_overruns);
1802
1803 static void rb_iter_reset(struct ring_buffer_iter *iter)
1804 {
1805         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
1806
1807         /* Iterator usage is expected to have record disabled */
1808         if (list_empty(&cpu_buffer->reader_page->list)) {
1809                 iter->head_page = cpu_buffer->head_page;
1810                 iter->head = cpu_buffer->head_page->read;
1811         } else {
1812                 iter->head_page = cpu_buffer->reader_page;
1813                 iter->head = cpu_buffer->reader_page->read;
1814         }
1815         if (iter->head)
1816                 iter->read_stamp = cpu_buffer->read_stamp;
1817         else
1818                 iter->read_stamp = iter->head_page->page->time_stamp;
1819 }
1820
1821 /**
1822  * ring_buffer_iter_reset - reset an iterator
1823  * @iter: The iterator to reset
1824  *
1825  * Resets the iterator, so that it will start from the beginning
1826  * again.
1827  */
1828 void ring_buffer_iter_reset(struct ring_buffer_iter *iter)
1829 {
1830         struct ring_buffer_per_cpu *cpu_buffer;
1831         unsigned long flags;
1832
1833         if (!iter)
1834                 return;
1835
1836         cpu_buffer = iter->cpu_buffer;
1837
1838         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
1839         rb_iter_reset(iter);
1840         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
1841 }
1842 EXPORT_SYMBOL_GPL(ring_buffer_iter_reset);
1843
1844 /**
1845  * ring_buffer_iter_empty - check if an iterator has no more to read
1846  * @iter: The iterator to check
1847  */
1848 int ring_buffer_iter_empty(struct ring_buffer_iter *iter)
1849 {
1850         struct ring_buffer_per_cpu *cpu_buffer;
1851
1852         cpu_buffer = iter->cpu_buffer;
1853
1854         return iter->head_page == cpu_buffer->commit_page &&
1855                 iter->head == rb_commit_index(cpu_buffer);
1856 }
1857 EXPORT_SYMBOL_GPL(ring_buffer_iter_empty);
1858
1859 static void
1860 rb_update_read_stamp(struct ring_buffer_per_cpu *cpu_buffer,
1861                      struct ring_buffer_event *event)
1862 {
1863         u64 delta;
1864
1865         switch (event->type) {
1866         case RINGBUF_TYPE_PADDING:
1867                 return;
1868
1869         case RINGBUF_TYPE_TIME_EXTEND:
1870                 delta = event->array[0];
1871                 delta <<= TS_SHIFT;
1872                 delta += event->time_delta;
1873                 cpu_buffer->read_stamp += delta;
1874                 return;
1875
1876         case RINGBUF_TYPE_TIME_STAMP:
1877                 /* FIXME: not implemented */
1878                 return;
1879
1880         case RINGBUF_TYPE_DATA:
1881                 cpu_buffer->read_stamp += event->time_delta;
1882                 return;
1883
1884         default:
1885                 BUG();
1886         }
1887         return;
1888 }
1889
1890 static void
1891 rb_update_iter_read_stamp(struct ring_buffer_iter *iter,
1892                           struct ring_buffer_event *event)
1893 {
1894         u64 delta;
1895
1896         switch (event->type) {
1897         case RINGBUF_TYPE_PADDING:
1898                 return;
1899
1900         case RINGBUF_TYPE_TIME_EXTEND:
1901                 delta = event->array[0];
1902                 delta <<= TS_SHIFT;
1903                 delta += event->time_delta;
1904                 iter->read_stamp += delta;
1905                 return;
1906
1907         case RINGBUF_TYPE_TIME_STAMP:
1908                 /* FIXME: not implemented */
1909                 return;
1910
1911         case RINGBUF_TYPE_DATA:
1912                 iter->read_stamp += event->time_delta;
1913                 return;
1914
1915         default:
1916                 BUG();
1917         }
1918         return;
1919 }
1920
1921 static struct buffer_page *
1922 rb_get_reader_page(struct ring_buffer_per_cpu *cpu_buffer)
1923 {
1924         struct buffer_page *reader = NULL;
1925         unsigned long flags;
1926         int nr_loops = 0;
1927
1928         local_irq_save(flags);
1929         __raw_spin_lock(&cpu_buffer->lock);
1930
1931  again:
1932         /*
1933          * This should normally only loop twice. But because the
1934          * start of the reader inserts an empty page, it causes
1935          * a case where we will loop three times. There should be no
1936          * reason to loop four times (that I know of).
1937          */
1938         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 3)) {
1939                 reader = NULL;
1940                 goto out;
1941         }
1942
1943         reader = cpu_buffer->reader_page;
1944
1945         /* If there's more to read, return this page */
1946         if (cpu_buffer->reader_page->read < rb_page_size(reader))
1947                 goto out;
1948
1949         /* Never should we have an index greater than the size */
1950         if (RB_WARN_ON(cpu_buffer,
1951                        cpu_buffer->reader_page->read > rb_page_size(reader)))
1952                 goto out;
1953
1954         /* check if we caught up to the tail */
1955         reader = NULL;
1956         if (cpu_buffer->commit_page == cpu_buffer->reader_page)
1957                 goto out;
1958
1959         /*
1960          * Splice the empty reader page into the list around the head.
1961          * Reset the reader page to size zero.
1962          */
1963
1964         reader = cpu_buffer->head_page;
1965         cpu_buffer->reader_page->list.next = reader->list.next;
1966         cpu_buffer->reader_page->list.prev = reader->list.prev;
1967
1968         local_set(&cpu_buffer->reader_page->write, 0);
1969         local_set(&cpu_buffer->reader_page->page->commit, 0);
1970
1971         /* Make the reader page now replace the head */
1972         reader->list.prev->next = &cpu_buffer->reader_page->list;
1973         reader->list.next->prev = &cpu_buffer->reader_page->list;
1974
1975         /*
1976          * If the tail is on the reader, then we must set the head
1977          * to the inserted page, otherwise we set it one before.
1978          */
1979         cpu_buffer->head_page = cpu_buffer->reader_page;
1980
1981         if (cpu_buffer->commit_page != reader)
1982                 rb_inc_page(cpu_buffer, &cpu_buffer->head_page);
1983
1984         /* Finally update the reader page to the new head */
1985         cpu_buffer->reader_page = reader;
1986         rb_reset_reader_page(cpu_buffer);
1987
1988         goto again;
1989
1990  out:
1991         __raw_spin_unlock(&cpu_buffer->lock);
1992         local_irq_restore(flags);
1993
1994         return reader;
1995 }
1996
1997 static void rb_advance_reader(struct ring_buffer_per_cpu *cpu_buffer)
1998 {
1999         struct ring_buffer_event *event;
2000         struct buffer_page *reader;
2001         unsigned length;
2002
2003         reader = rb_get_reader_page(cpu_buffer);
2004
2005         /* This function should not be called when buffer is empty */
2006         if (RB_WARN_ON(cpu_buffer, !reader))
2007                 return;
2008
2009         event = rb_reader_event(cpu_buffer);
2010
2011         if (event->type == RINGBUF_TYPE_DATA || rb_discarded_event(event))
2012                 cpu_buffer->entries--;
2013
2014         rb_update_read_stamp(cpu_buffer, event);
2015
2016         length = rb_event_length(event);
2017         cpu_buffer->reader_page->read += length;
2018 }
2019
2020 static void rb_advance_iter(struct ring_buffer_iter *iter)
2021 {
2022         struct ring_buffer *buffer;
2023         struct ring_buffer_per_cpu *cpu_buffer;
2024         struct ring_buffer_event *event;
2025         unsigned length;
2026
2027         cpu_buffer = iter->cpu_buffer;
2028         buffer = cpu_buffer->buffer;
2029
2030         /*
2031          * Check if we are at the end of the buffer.
2032          */
2033         if (iter->head >= rb_page_size(iter->head_page)) {
2034                 if (RB_WARN_ON(buffer,
2035                                iter->head_page == cpu_buffer->commit_page))
2036                         return;
2037                 rb_inc_iter(iter);
2038                 return;
2039         }
2040
2041         event = rb_iter_head_event(iter);
2042
2043         length = rb_event_length(event);
2044
2045         /*
2046          * This should not be called to advance the header if we are
2047          * at the tail of the buffer.
2048          */
2049         if (RB_WARN_ON(cpu_buffer,
2050                        (iter->head_page == cpu_buffer->commit_page) &&
2051                        (iter->head + length > rb_commit_index(cpu_buffer))))
2052                 return;
2053
2054         rb_update_iter_read_stamp(iter, event);
2055
2056         iter->head += length;
2057
2058         /* check for end of page padding */
2059         if ((iter->head >= rb_page_size(iter->head_page)) &&
2060             (iter->head_page != cpu_buffer->commit_page))
2061                 rb_advance_iter(iter);
2062 }
2063
2064 static struct ring_buffer_event *
2065 rb_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts)
2066 {
2067         struct ring_buffer_per_cpu *cpu_buffer;
2068         struct ring_buffer_event *event;
2069         struct buffer_page *reader;
2070         int nr_loops = 0;
2071
2072         cpu_buffer = buffer->buffers[cpu];
2073
2074  again:
2075         /*
2076          * We repeat when a timestamp is encountered. It is possible
2077          * to get multiple timestamps from an interrupt entering just
2078          * as one timestamp is about to be written. The max times
2079          * that this can happen is the number of nested interrupts we
2080          * can have.  Nesting 10 deep of interrupts is clearly
2081          * an anomaly.
2082          */
2083         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 10))
2084                 return NULL;
2085
2086         reader = rb_get_reader_page(cpu_buffer);
2087         if (!reader)
2088                 return NULL;
2089
2090         event = rb_reader_event(cpu_buffer);
2091
2092         switch (event->type) {
2093         case RINGBUF_TYPE_PADDING:
2094                 if (rb_null_event(event))
2095                         RB_WARN_ON(cpu_buffer, 1);
2096                 /*
2097                  * Because the writer could be discarding every
2098                  * event it creates (which would probably be bad)
2099                  * if we were to go back to "again" then we may never
2100                  * catch up, and will trigger the warn on, or lock
2101                  * the box. Return the padding, and we will release
2102                  * the current locks, and try again.
2103                  */
2104                 rb_advance_reader(cpu_buffer);
2105                 return event;
2106
2107         case RINGBUF_TYPE_TIME_EXTEND:
2108                 /* Internal data, OK to advance */
2109                 rb_advance_reader(cpu_buffer);
2110                 goto again;
2111
2112         case RINGBUF_TYPE_TIME_STAMP:
2113                 /* FIXME: not implemented */
2114                 rb_advance_reader(cpu_buffer);
2115                 goto again;
2116
2117         case RINGBUF_TYPE_DATA:
2118                 if (ts) {
2119                         *ts = cpu_buffer->read_stamp + event->time_delta;
2120                         ring_buffer_normalize_time_stamp(buffer,
2121                                                          cpu_buffer->cpu, ts);
2122                 }
2123                 return event;
2124
2125         default:
2126                 BUG();
2127         }
2128
2129         return NULL;
2130 }
2131 EXPORT_SYMBOL_GPL(ring_buffer_peek);
2132
2133 static struct ring_buffer_event *
2134 rb_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
2135 {
2136         struct ring_buffer *buffer;
2137         struct ring_buffer_per_cpu *cpu_buffer;
2138         struct ring_buffer_event *event;
2139         int nr_loops = 0;
2140
2141         if (ring_buffer_iter_empty(iter))
2142                 return NULL;
2143
2144         cpu_buffer = iter->cpu_buffer;
2145         buffer = cpu_buffer->buffer;
2146
2147  again:
2148         /*
2149          * We repeat when a timestamp is encountered. It is possible
2150          * to get multiple timestamps from an interrupt entering just
2151          * as one timestamp is about to be written. The max times
2152          * that this can happen is the number of nested interrupts we
2153          * can have. Nesting 10 deep of interrupts is clearly
2154          * an anomaly.
2155          */
2156         if (RB_WARN_ON(cpu_buffer, ++nr_loops > 10))
2157                 return NULL;
2158
2159         if (rb_per_cpu_empty(cpu_buffer))
2160                 return NULL;
2161
2162         event = rb_iter_head_event(iter);
2163
2164         switch (event->type) {
2165         case RINGBUF_TYPE_PADDING:
2166                 if (rb_null_event(event)) {
2167                         rb_inc_iter(iter);
2168                         goto again;
2169                 }
2170                 rb_advance_iter(iter);
2171                 return event;
2172
2173         case RINGBUF_TYPE_TIME_EXTEND:
2174                 /* Internal data, OK to advance */
2175                 rb_advance_iter(iter);
2176                 goto again;
2177
2178         case RINGBUF_TYPE_TIME_STAMP:
2179                 /* FIXME: not implemented */
2180                 rb_advance_iter(iter);
2181                 goto again;
2182
2183         case RINGBUF_TYPE_DATA:
2184                 if (ts) {
2185                         *ts = iter->read_stamp + event->time_delta;
2186                         ring_buffer_normalize_time_stamp(buffer,
2187                                                          cpu_buffer->cpu, ts);
2188                 }
2189                 return event;
2190
2191         default:
2192                 BUG();
2193         }
2194
2195         return NULL;
2196 }
2197 EXPORT_SYMBOL_GPL(ring_buffer_iter_peek);
2198
2199 /**
2200  * ring_buffer_peek - peek at the next event to be read
2201  * @buffer: The ring buffer to read
2202  * @cpu: The cpu to peak at
2203  * @ts: The timestamp counter of this event.
2204  *
2205  * This will return the event that will be read next, but does
2206  * not consume the data.
2207  */
2208 struct ring_buffer_event *
2209 ring_buffer_peek(struct ring_buffer *buffer, int cpu, u64 *ts)
2210 {
2211         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
2212         struct ring_buffer_event *event;
2213         unsigned long flags;
2214
2215         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2216                 return NULL;
2217
2218  again:
2219         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2220         event = rb_buffer_peek(buffer, cpu, ts);
2221         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2222
2223         if (event && event->type == RINGBUF_TYPE_PADDING) {
2224                 cpu_relax();
2225                 goto again;
2226         }
2227
2228         return event;
2229 }
2230
2231 /**
2232  * ring_buffer_iter_peek - peek at the next event to be read
2233  * @iter: The ring buffer iterator
2234  * @ts: The timestamp counter of this event.
2235  *
2236  * This will return the event that will be read next, but does
2237  * not increment the iterator.
2238  */
2239 struct ring_buffer_event *
2240 ring_buffer_iter_peek(struct ring_buffer_iter *iter, u64 *ts)
2241 {
2242         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2243         struct ring_buffer_event *event;
2244         unsigned long flags;
2245
2246  again:
2247         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2248         event = rb_iter_peek(iter, ts);
2249         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2250
2251         if (event && event->type == RINGBUF_TYPE_PADDING) {
2252                 cpu_relax();
2253                 goto again;
2254         }
2255
2256         return event;
2257 }
2258
2259 /**
2260  * ring_buffer_consume - return an event and consume it
2261  * @buffer: The ring buffer to get the next event from
2262  *
2263  * Returns the next event in the ring buffer, and that event is consumed.
2264  * Meaning, that sequential reads will keep returning a different event,
2265  * and eventually empty the ring buffer if the producer is slower.
2266  */
2267 struct ring_buffer_event *
2268 ring_buffer_consume(struct ring_buffer *buffer, int cpu, u64 *ts)
2269 {
2270         struct ring_buffer_per_cpu *cpu_buffer;
2271         struct ring_buffer_event *event = NULL;
2272         unsigned long flags;
2273
2274  again:
2275         /* might be called in atomic */
2276         preempt_disable();
2277
2278         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2279                 goto out;
2280
2281         cpu_buffer = buffer->buffers[cpu];
2282         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2283
2284         event = rb_buffer_peek(buffer, cpu, ts);
2285         if (!event)
2286                 goto out_unlock;
2287
2288         rb_advance_reader(cpu_buffer);
2289
2290  out_unlock:
2291         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2292
2293  out:
2294         preempt_enable();
2295
2296         if (event && event->type == RINGBUF_TYPE_PADDING) {
2297                 cpu_relax();
2298                 goto again;
2299         }
2300
2301         return event;
2302 }
2303 EXPORT_SYMBOL_GPL(ring_buffer_consume);
2304
2305 /**
2306  * ring_buffer_read_start - start a non consuming read of the buffer
2307  * @buffer: The ring buffer to read from
2308  * @cpu: The cpu buffer to iterate over
2309  *
2310  * This starts up an iteration through the buffer. It also disables
2311  * the recording to the buffer until the reading is finished.
2312  * This prevents the reading from being corrupted. This is not
2313  * a consuming read, so a producer is not expected.
2314  *
2315  * Must be paired with ring_buffer_finish.
2316  */
2317 struct ring_buffer_iter *
2318 ring_buffer_read_start(struct ring_buffer *buffer, int cpu)
2319 {
2320         struct ring_buffer_per_cpu *cpu_buffer;
2321         struct ring_buffer_iter *iter;
2322         unsigned long flags;
2323
2324         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2325                 return NULL;
2326
2327         iter = kmalloc(sizeof(*iter), GFP_KERNEL);
2328         if (!iter)
2329                 return NULL;
2330
2331         cpu_buffer = buffer->buffers[cpu];
2332
2333         iter->cpu_buffer = cpu_buffer;
2334
2335         atomic_inc(&cpu_buffer->record_disabled);
2336         synchronize_sched();
2337
2338         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2339         __raw_spin_lock(&cpu_buffer->lock);
2340         rb_iter_reset(iter);
2341         __raw_spin_unlock(&cpu_buffer->lock);
2342         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2343
2344         return iter;
2345 }
2346 EXPORT_SYMBOL_GPL(ring_buffer_read_start);
2347
2348 /**
2349  * ring_buffer_finish - finish reading the iterator of the buffer
2350  * @iter: The iterator retrieved by ring_buffer_start
2351  *
2352  * This re-enables the recording to the buffer, and frees the
2353  * iterator.
2354  */
2355 void
2356 ring_buffer_read_finish(struct ring_buffer_iter *iter)
2357 {
2358         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2359
2360         atomic_dec(&cpu_buffer->record_disabled);
2361         kfree(iter);
2362 }
2363 EXPORT_SYMBOL_GPL(ring_buffer_read_finish);
2364
2365 /**
2366  * ring_buffer_read - read the next item in the ring buffer by the iterator
2367  * @iter: The ring buffer iterator
2368  * @ts: The time stamp of the event read.
2369  *
2370  * This reads the next event in the ring buffer and increments the iterator.
2371  */
2372 struct ring_buffer_event *
2373 ring_buffer_read(struct ring_buffer_iter *iter, u64 *ts)
2374 {
2375         struct ring_buffer_event *event;
2376         struct ring_buffer_per_cpu *cpu_buffer = iter->cpu_buffer;
2377         unsigned long flags;
2378
2379  again:
2380         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2381         event = rb_iter_peek(iter, ts);
2382         if (!event)
2383                 goto out;
2384
2385         rb_advance_iter(iter);
2386  out:
2387         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2388
2389         if (event && event->type == RINGBUF_TYPE_PADDING) {
2390                 cpu_relax();
2391                 goto again;
2392         }
2393
2394         return event;
2395 }
2396 EXPORT_SYMBOL_GPL(ring_buffer_read);
2397
2398 /**
2399  * ring_buffer_size - return the size of the ring buffer (in bytes)
2400  * @buffer: The ring buffer.
2401  */
2402 unsigned long ring_buffer_size(struct ring_buffer *buffer)
2403 {
2404         return BUF_PAGE_SIZE * buffer->pages;
2405 }
2406 EXPORT_SYMBOL_GPL(ring_buffer_size);
2407
2408 static void
2409 rb_reset_cpu(struct ring_buffer_per_cpu *cpu_buffer)
2410 {
2411         cpu_buffer->head_page
2412                 = list_entry(cpu_buffer->pages.next, struct buffer_page, list);
2413         local_set(&cpu_buffer->head_page->write, 0);
2414         local_set(&cpu_buffer->head_page->page->commit, 0);
2415
2416         cpu_buffer->head_page->read = 0;
2417
2418         cpu_buffer->tail_page = cpu_buffer->head_page;
2419         cpu_buffer->commit_page = cpu_buffer->head_page;
2420
2421         INIT_LIST_HEAD(&cpu_buffer->reader_page->list);
2422         local_set(&cpu_buffer->reader_page->write, 0);
2423         local_set(&cpu_buffer->reader_page->page->commit, 0);
2424         cpu_buffer->reader_page->read = 0;
2425
2426         cpu_buffer->overrun = 0;
2427         cpu_buffer->entries = 0;
2428
2429         cpu_buffer->write_stamp = 0;
2430         cpu_buffer->read_stamp = 0;
2431 }
2432
2433 /**
2434  * ring_buffer_reset_cpu - reset a ring buffer per CPU buffer
2435  * @buffer: The ring buffer to reset a per cpu buffer of
2436  * @cpu: The CPU buffer to be reset
2437  */
2438 void ring_buffer_reset_cpu(struct ring_buffer *buffer, int cpu)
2439 {
2440         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
2441         unsigned long flags;
2442
2443         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2444                 return;
2445
2446         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2447
2448         __raw_spin_lock(&cpu_buffer->lock);
2449
2450         rb_reset_cpu(cpu_buffer);
2451
2452         __raw_spin_unlock(&cpu_buffer->lock);
2453
2454         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2455 }
2456 EXPORT_SYMBOL_GPL(ring_buffer_reset_cpu);
2457
2458 /**
2459  * ring_buffer_reset - reset a ring buffer
2460  * @buffer: The ring buffer to reset all cpu buffers
2461  */
2462 void ring_buffer_reset(struct ring_buffer *buffer)
2463 {
2464         int cpu;
2465
2466         for_each_buffer_cpu(buffer, cpu)
2467                 ring_buffer_reset_cpu(buffer, cpu);
2468 }
2469 EXPORT_SYMBOL_GPL(ring_buffer_reset);
2470
2471 /**
2472  * rind_buffer_empty - is the ring buffer empty?
2473  * @buffer: The ring buffer to test
2474  */
2475 int ring_buffer_empty(struct ring_buffer *buffer)
2476 {
2477         struct ring_buffer_per_cpu *cpu_buffer;
2478         int cpu;
2479
2480         /* yes this is racy, but if you don't like the race, lock the buffer */
2481         for_each_buffer_cpu(buffer, cpu) {
2482                 cpu_buffer = buffer->buffers[cpu];
2483                 if (!rb_per_cpu_empty(cpu_buffer))
2484                         return 0;
2485         }
2486
2487         return 1;
2488 }
2489 EXPORT_SYMBOL_GPL(ring_buffer_empty);
2490
2491 /**
2492  * ring_buffer_empty_cpu - is a cpu buffer of a ring buffer empty?
2493  * @buffer: The ring buffer
2494  * @cpu: The CPU buffer to test
2495  */
2496 int ring_buffer_empty_cpu(struct ring_buffer *buffer, int cpu)
2497 {
2498         struct ring_buffer_per_cpu *cpu_buffer;
2499         int ret;
2500
2501         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2502                 return 1;
2503
2504         cpu_buffer = buffer->buffers[cpu];
2505         ret = rb_per_cpu_empty(cpu_buffer);
2506
2507
2508         return ret;
2509 }
2510 EXPORT_SYMBOL_GPL(ring_buffer_empty_cpu);
2511
2512 /**
2513  * ring_buffer_swap_cpu - swap a CPU buffer between two ring buffers
2514  * @buffer_a: One buffer to swap with
2515  * @buffer_b: The other buffer to swap with
2516  *
2517  * This function is useful for tracers that want to take a "snapshot"
2518  * of a CPU buffer and has another back up buffer lying around.
2519  * it is expected that the tracer handles the cpu buffer not being
2520  * used at the moment.
2521  */
2522 int ring_buffer_swap_cpu(struct ring_buffer *buffer_a,
2523                          struct ring_buffer *buffer_b, int cpu)
2524 {
2525         struct ring_buffer_per_cpu *cpu_buffer_a;
2526         struct ring_buffer_per_cpu *cpu_buffer_b;
2527         int ret = -EINVAL;
2528
2529         if (!cpumask_test_cpu(cpu, buffer_a->cpumask) ||
2530             !cpumask_test_cpu(cpu, buffer_b->cpumask))
2531                 goto out;
2532
2533         /* At least make sure the two buffers are somewhat the same */
2534         if (buffer_a->pages != buffer_b->pages)
2535                 goto out;
2536
2537         ret = -EAGAIN;
2538
2539         if (ring_buffer_flags != RB_BUFFERS_ON)
2540                 goto out;
2541
2542         if (atomic_read(&buffer_a->record_disabled))
2543                 goto out;
2544
2545         if (atomic_read(&buffer_b->record_disabled))
2546                 goto out;
2547
2548         cpu_buffer_a = buffer_a->buffers[cpu];
2549         cpu_buffer_b = buffer_b->buffers[cpu];
2550
2551         if (atomic_read(&cpu_buffer_a->record_disabled))
2552                 goto out;
2553
2554         if (atomic_read(&cpu_buffer_b->record_disabled))
2555                 goto out;
2556
2557         /*
2558          * We can't do a synchronize_sched here because this
2559          * function can be called in atomic context.
2560          * Normally this will be called from the same CPU as cpu.
2561          * If not it's up to the caller to protect this.
2562          */
2563         atomic_inc(&cpu_buffer_a->record_disabled);
2564         atomic_inc(&cpu_buffer_b->record_disabled);
2565
2566         buffer_a->buffers[cpu] = cpu_buffer_b;
2567         buffer_b->buffers[cpu] = cpu_buffer_a;
2568
2569         cpu_buffer_b->buffer = buffer_a;
2570         cpu_buffer_a->buffer = buffer_b;
2571
2572         atomic_dec(&cpu_buffer_a->record_disabled);
2573         atomic_dec(&cpu_buffer_b->record_disabled);
2574
2575         ret = 0;
2576 out:
2577         return ret;
2578 }
2579 EXPORT_SYMBOL_GPL(ring_buffer_swap_cpu);
2580
2581 static void rb_remove_entries(struct ring_buffer_per_cpu *cpu_buffer,
2582                               struct buffer_data_page *bpage,
2583                               unsigned int offset)
2584 {
2585         struct ring_buffer_event *event;
2586         unsigned long head;
2587
2588         __raw_spin_lock(&cpu_buffer->lock);
2589         for (head = offset; head < local_read(&bpage->commit);
2590              head += rb_event_length(event)) {
2591
2592                 event = __rb_data_page_index(bpage, head);
2593                 if (RB_WARN_ON(cpu_buffer, rb_null_event(event)))
2594                         return;
2595                 /* Only count data entries */
2596                 if (event->type != RINGBUF_TYPE_DATA)
2597                         continue;
2598                 cpu_buffer->entries--;
2599         }
2600         __raw_spin_unlock(&cpu_buffer->lock);
2601 }
2602
2603 /**
2604  * ring_buffer_alloc_read_page - allocate a page to read from buffer
2605  * @buffer: the buffer to allocate for.
2606  *
2607  * This function is used in conjunction with ring_buffer_read_page.
2608  * When reading a full page from the ring buffer, these functions
2609  * can be used to speed up the process. The calling function should
2610  * allocate a few pages first with this function. Then when it
2611  * needs to get pages from the ring buffer, it passes the result
2612  * of this function into ring_buffer_read_page, which will swap
2613  * the page that was allocated, with the read page of the buffer.
2614  *
2615  * Returns:
2616  *  The page allocated, or NULL on error.
2617  */
2618 void *ring_buffer_alloc_read_page(struct ring_buffer *buffer)
2619 {
2620         struct buffer_data_page *bpage;
2621         unsigned long addr;
2622
2623         addr = __get_free_page(GFP_KERNEL);
2624         if (!addr)
2625                 return NULL;
2626
2627         bpage = (void *)addr;
2628
2629         rb_init_page(bpage);
2630
2631         return bpage;
2632 }
2633
2634 /**
2635  * ring_buffer_free_read_page - free an allocated read page
2636  * @buffer: the buffer the page was allocate for
2637  * @data: the page to free
2638  *
2639  * Free a page allocated from ring_buffer_alloc_read_page.
2640  */
2641 void ring_buffer_free_read_page(struct ring_buffer *buffer, void *data)
2642 {
2643         free_page((unsigned long)data);
2644 }
2645
2646 /**
2647  * ring_buffer_read_page - extract a page from the ring buffer
2648  * @buffer: buffer to extract from
2649  * @data_page: the page to use allocated from ring_buffer_alloc_read_page
2650  * @len: amount to extract
2651  * @cpu: the cpu of the buffer to extract
2652  * @full: should the extraction only happen when the page is full.
2653  *
2654  * This function will pull out a page from the ring buffer and consume it.
2655  * @data_page must be the address of the variable that was returned
2656  * from ring_buffer_alloc_read_page. This is because the page might be used
2657  * to swap with a page in the ring buffer.
2658  *
2659  * for example:
2660  *      rpage = ring_buffer_alloc_read_page(buffer);
2661  *      if (!rpage)
2662  *              return error;
2663  *      ret = ring_buffer_read_page(buffer, &rpage, len, cpu, 0);
2664  *      if (ret >= 0)
2665  *              process_page(rpage, ret);
2666  *
2667  * When @full is set, the function will not return true unless
2668  * the writer is off the reader page.
2669  *
2670  * Note: it is up to the calling functions to handle sleeps and wakeups.
2671  *  The ring buffer can be used anywhere in the kernel and can not
2672  *  blindly call wake_up. The layer that uses the ring buffer must be
2673  *  responsible for that.
2674  *
2675  * Returns:
2676  *  >=0 if data has been transferred, returns the offset of consumed data.
2677  *  <0 if no data has been transferred.
2678  */
2679 int ring_buffer_read_page(struct ring_buffer *buffer,
2680                           void **data_page, size_t len, int cpu, int full)
2681 {
2682         struct ring_buffer_per_cpu *cpu_buffer = buffer->buffers[cpu];
2683         struct ring_buffer_event *event;
2684         struct buffer_data_page *bpage;
2685         struct buffer_page *reader;
2686         unsigned long flags;
2687         unsigned int commit;
2688         unsigned int read;
2689         u64 save_timestamp;
2690         int ret = -1;
2691
2692         if (!cpumask_test_cpu(cpu, buffer->cpumask))
2693                 goto out;
2694
2695         /*
2696          * If len is not big enough to hold the page header, then
2697          * we can not copy anything.
2698          */
2699         if (len <= BUF_PAGE_HDR_SIZE)
2700                 goto out;
2701
2702         len -= BUF_PAGE_HDR_SIZE;
2703
2704         if (!data_page)
2705                 goto out;
2706
2707         bpage = *data_page;
2708         if (!bpage)
2709                 goto out;
2710
2711         spin_lock_irqsave(&cpu_buffer->reader_lock, flags);
2712
2713         reader = rb_get_reader_page(cpu_buffer);
2714         if (!reader)
2715                 goto out_unlock;
2716
2717         event = rb_reader_event(cpu_buffer);
2718
2719         read = reader->read;
2720         commit = rb_page_commit(reader);
2721
2722         /*
2723          * If this page has been partially read or
2724          * if len is not big enough to read the rest of the page or
2725          * a writer is still on the page, then
2726          * we must copy the data from the page to the buffer.
2727          * Otherwise, we can simply swap the page with the one passed in.
2728          */
2729         if (read || (len < (commit - read)) ||
2730             cpu_buffer->reader_page == cpu_buffer->commit_page) {
2731                 struct buffer_data_page *rpage = cpu_buffer->reader_page->page;
2732                 unsigned int rpos = read;
2733                 unsigned int pos = 0;
2734                 unsigned int size;
2735
2736                 if (full)
2737                         goto out_unlock;
2738
2739                 if (len > (commit - read))
2740                         len = (commit - read);
2741
2742                 size = rb_event_length(event);
2743
2744                 if (len < size)
2745                         goto out_unlock;
2746
2747                 /* save the current timestamp, since the user will need it */
2748                 save_timestamp = cpu_buffer->read_stamp;
2749
2750                 /* Need to copy one event at a time */
2751                 do {
2752                         memcpy(bpage->data + pos, rpage->data + rpos, size);
2753
2754                         len -= size;
2755
2756                         rb_advance_reader(cpu_buffer);
2757                         rpos = reader->read;
2758                         pos += size;
2759
2760                         event = rb_reader_event(cpu_buffer);
2761                         size = rb_event_length(event);
2762                 } while (len > size);
2763
2764                 /* update bpage */
2765                 local_set(&bpage->commit, pos);
2766                 bpage->time_stamp = save_timestamp;
2767
2768                 /* we copied everything to the beginning */
2769                 read = 0;
2770         } else {
2771                 /* swap the pages */
2772                 rb_init_page(bpage);
2773                 bpage = reader->page;
2774                 reader->page = *data_page;
2775                 local_set(&reader->write, 0);
2776                 reader->read = 0;
2777                 *data_page = bpage;
2778
2779                 /* update the entry counter */
2780                 rb_remove_entries(cpu_buffer, bpage, read);
2781         }
2782         ret = read;
2783
2784  out_unlock:
2785         spin_unlock_irqrestore(&cpu_buffer->reader_lock, flags);
2786
2787  out:
2788         return ret;
2789 }
2790
2791 static ssize_t
2792 rb_simple_read(struct file *filp, char __user *ubuf,
2793                size_t cnt, loff_t *ppos)
2794 {
2795         unsigned long *p = filp->private_data;
2796         char buf[64];
2797         int r;
2798
2799         if (test_bit(RB_BUFFERS_DISABLED_BIT, p))
2800                 r = sprintf(buf, "permanently disabled\n");
2801         else
2802                 r = sprintf(buf, "%d\n", test_bit(RB_BUFFERS_ON_BIT, p));
2803
2804         return simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
2805 }
2806
2807 static ssize_t
2808 rb_simple_write(struct file *filp, const char __user *ubuf,
2809                 size_t cnt, loff_t *ppos)
2810 {
2811         unsigned long *p = filp->private_data;
2812         char buf[64];
2813         unsigned long val;
2814         int ret;
2815
2816         if (cnt >= sizeof(buf))
2817                 return -EINVAL;
2818
2819         if (copy_from_user(&buf, ubuf, cnt))
2820                 return -EFAULT;
2821
2822         buf[cnt] = 0;
2823
2824         ret = strict_strtoul(buf, 10, &val);
2825         if (ret < 0)
2826                 return ret;
2827
2828         if (val)
2829                 set_bit(RB_BUFFERS_ON_BIT, p);
2830         else
2831                 clear_bit(RB_BUFFERS_ON_BIT, p);
2832
2833         (*ppos)++;
2834
2835         return cnt;
2836 }
2837
2838 static const struct file_operations rb_simple_fops = {
2839         .open           = tracing_open_generic,
2840         .read           = rb_simple_read,
2841         .write          = rb_simple_write,
2842 };
2843
2844
2845 static __init int rb_init_debugfs(void)
2846 {
2847         struct dentry *d_tracer;
2848         struct dentry *entry;
2849
2850         d_tracer = tracing_init_dentry();
2851
2852         entry = debugfs_create_file("tracing_on", 0644, d_tracer,
2853                                     &ring_buffer_flags, &rb_simple_fops);
2854         if (!entry)
2855                 pr_warning("Could not create debugfs 'tracing_on' entry\n");
2856
2857         return 0;
2858 }
2859
2860 fs_initcall(rb_init_debugfs);
2861
2862 #ifdef CONFIG_HOTPLUG_CPU
2863 static int rb_cpu_notify(struct notifier_block *self,
2864                          unsigned long action, void *hcpu)
2865 {
2866         struct ring_buffer *buffer =
2867                 container_of(self, struct ring_buffer, cpu_notify);
2868         long cpu = (long)hcpu;
2869
2870         switch (action) {
2871         case CPU_UP_PREPARE:
2872         case CPU_UP_PREPARE_FROZEN:
2873                 if (cpu_isset(cpu, *buffer->cpumask))
2874                         return NOTIFY_OK;
2875
2876                 buffer->buffers[cpu] =
2877                         rb_allocate_cpu_buffer(buffer, cpu);
2878                 if (!buffer->buffers[cpu]) {
2879                         WARN(1, "failed to allocate ring buffer on CPU %ld\n",
2880                              cpu);
2881                         return NOTIFY_OK;
2882                 }
2883                 smp_wmb();
2884                 cpu_set(cpu, *buffer->cpumask);
2885                 break;
2886         case CPU_DOWN_PREPARE:
2887         case CPU_DOWN_PREPARE_FROZEN:
2888                 /*
2889                  * Do nothing.
2890                  *  If we were to free the buffer, then the user would
2891                  *  lose any trace that was in the buffer.
2892                  */
2893                 break;
2894         default:
2895                 break;
2896         }
2897         return NOTIFY_OK;
2898 }
2899 #endif